A method and system for online anomaly monitoring of a linear moving cutting slurry sampler

Abstract

This invention discloses an online anomaly monitoring method and system for a linear moving cutting slurry sampler, relating to the field of industrial automation technology. The method includes performing frequency band energy separation and sliding window statistical analysis on the slurry sampler's operating condition dataset to obtain a multi-dimensional feature matrix; performing weight allocation and dynamic weighted aggregation on the multi-dimensional feature matrix to form a spatiotemporal analysis data package; performing synergy analysis on the spatiotemporal analysis data package to output a trend synergy interaction matrix; and using the entropy weight method to perform risk quantification and contribution allocation on the trend synergy interaction matrix to generate anomaly quantification parameters. The anomaly quantification parameters are then weighted with confidence to form anomaly probability values. This invention fully integrates the slurry sampler's operating condition dataset through sliding window statistical analysis and the entropy weight method, while simultaneously improving the reliability of anomaly monitoring through deep feature mining and spatial relationship fusion using dynamic causal graphs and spatiotemporal convolutional neural network models.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, and in particular to an online anomaly monitoring method and system for a linear moving cutting slurry sampler. Background Technology

[0002] With the continuous improvement of industrial automation and intelligence, linear moving cutting slurry samplers, as one of the key pieces of equipment in the mining production process, have received increasing attention for monitoring and maintaining their operating status. Traditional monitoring methods mainly rely on manual inspections and periodic maintenance. In recent years, with the rapid development of sensing technology, data acquisition systems, and edge computing devices, intelligent operation and maintenance methods based on online monitoring and data analysis have been gradually applied to the status perception and fault early warning of slurry samplers.

[0003] Despite the progress made in existing monitoring methods, some shortcomings remain. First, when dealing with the fusion of multi-source heterogeneous data, traditional monitoring methods typically employ simple weighted averaging or direct data concatenation, leading to the loss of important information and the neglect of relationships between features. Second, traditional anomaly monitoring systems often rely on shallow classification models or LSTM sequence models, which struggle to characterize the dynamic behavior of slurry samplers under complex operating conditions and the intricate causal relationships between components, thus reducing the reliability of anomaly detection. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an online anomaly monitoring method for a linear moving cutting slurry sampler to address the problems of insufficient multi-source data fusion and inadequate dynamic causal relationship modeling.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an online anomaly monitoring method for a linear moving cutting slurry sampler, comprising performing frequency band energy separation and sliding window statistical analysis on the slurry sampler operating condition dataset to obtain a multidimensional feature matrix, and performing weight allocation and dynamic weighted aggregation on the multidimensional feature matrix to form a spatiotemporal analysis data package;

[0008] Perform synergy analysis on the spatiotemporal analysis data package to output a trend synergy interaction matrix; use the entropy weight method to perform risk quantification and contribution allocation on the trend synergy interaction matrix to generate anomaly quantification parameters, and perform confidence weighted calculation on the anomaly quantification parameters to form anomaly probability values; use the time decay compensation factor to perform time-series correction and spatial correlation mapping on the anomaly probability values ​​to obtain a spatiotemporal anomaly index vector.

[0009] The spatiotemporal anomaly index vector is analyzed for correlation strength and decomposed by singular value, state features and correlation features are extracted, and the state features are mapped to nodes of the topological structure and the correlation features are mapped to edges of the topological structure to construct a dynamic causal graph.

[0010] The dynamic causal graph is input into the spatiotemporal convolutional neural network model. The feature aggregation layer uses the adjacency normalization operator to perform node feature extraction and spatial relationship fusion. The spatiotemporal coupling layer performs multi-scale feature enhancement through the residual cross-layer connection mechanism to form an anomaly impact score. The anomaly impact score is then classified and structured to output an anomaly monitoring report.

[0011] As a preferred embodiment of the online anomaly monitoring method for the linear moving cut slurry sampler described in this invention, the slurry sampler operating condition dataset includes mechanical moving cut parameters, slurry physical property parameters, and slurry sampler electrical parameters.

[0012] As a preferred embodiment of the online anomaly monitoring method for the linear moving cutting slurry sampler described in this invention, the formation of the spatiotemporal analysis data package specifically includes the following steps.

[0013] Fast Fourier Transform is applied to perform frequency band energy separation on the slurry sampler operating condition dataset to form a time-frequency composite parameter sequence; sliding window statistical analysis is performed on the time-frequency composite parameter sequence to obtain a multidimensional feature matrix;

[0014] Weighting and vector reconstruction are performed on the multidimensional feature matrix to generate a weighted spatiotemporal vector; dynamic weighted aggregation is performed on the weighted spatiotemporal vector to output a spatiotemporal analysis data package.

[0015] As a preferred embodiment of the online anomaly monitoring method for the linear moving cutting slurry sampler described in this invention, the step of acquiring the spatiotemporal anomaly index vector specifically includes the following steps.

[0016] The time decay compensation factor is used to perform window smoothing and time weight allocation on the abnormal probability value to generate a decay correction sequence. The abnormal probability value is then time-series corrected according to the decay correction sequence to generate a time-series correction vector.

[0017] The temporal correction vector is divided into regions and features are aggregated to obtain a spatial feature set. The spatial feature set is then spatially correlated and mapped to generate a spatiotemporal anomaly index vector.

[0018] As a preferred embodiment of the online anomaly monitoring method for the linear moving cutting slurry sampler described in this invention, the construction of the dynamic causal map specifically includes the following steps.

[0019] The correlation strength of the spatiotemporal anomaly index vector is analyzed to form a correlation strength matrix. Singular value decomposition is then performed on the correlation strength matrix to extract state features and correlation features.

[0020] The state features are normalized in dimension to obtain standard state parameters. Node attribute mapping is then performed on the standard state parameters to generate nodes in the topology structure.

[0021] The interaction of associated features is quantified to generate an edge weight matrix. The edge strength of the edge weight matrix is ​​then mapped to obtain the edges of the topological structure.

[0022] Dynamic hierarchical clustering is performed on the nodes and edges of the topology to construct a dynamic causal graph.

[0023] As a preferred embodiment of the online anomaly monitoring method for the linear moving cutting slurry sampler described in this invention, the spatiotemporal convolutional neural network model is constructed by building a feature aggregation layer and a spatiotemporal coupling layer, and applying a multi-head attention mechanism for cross-layer parameterization stacking.

[0024] As a preferred embodiment of the online anomaly monitoring method for the linear moving cutting slurry sampler described in this invention, the formation of an anomaly impact score specifically includes the following steps.

[0025] The dynamic causal graph is input into the spatiotemporal convolutional neural network model. The feature aggregation layer applies the adjacency normalization operator to perform node feature extraction and spatial relationship fusion to generate a primary node embedding representation.

[0026] The spatiotemporal coupling layer enhances multi-scale features through a residual cross-layer connection mechanism to obtain an enhanced spatiotemporal feature matrix;

[0027] The primary node embedding representation and the enhanced spatiotemporal feature matrix are subjected to fully connected transformation and probability normalization to form an anomaly impact score.

[0028] Secondly, the present invention provides an online anomaly monitoring system for a linear moving cutting slurry sampler, comprising,

[0029] The data analysis module is used to perform frequency band energy separation and sliding window statistical analysis on the slurry sampler operating condition dataset, obtain a multidimensional feature matrix, and perform weight allocation and dynamic weighted aggregation on the multidimensional feature matrix to form a spatiotemporal analysis data package;

[0030] The anomaly quantification module is used to perform collaborative analysis on spatiotemporal analysis data packets and output a trend collaborative interaction matrix. It uses the entropy weight method to perform risk quantification and contribution allocation on the trend collaborative interaction matrix, generates anomaly quantification parameters, and performs confidence weighted calculation on the anomaly quantification parameters to form anomaly probability values. It uses the time decay compensation factor to perform time-series correction and spatial correlation mapping on the anomaly probability values ​​to obtain a spatiotemporal anomaly index vector.

[0031] The graph construction module is used to perform correlation strength analysis and singular value decomposition on spatiotemporal anomaly index vectors, extract state features and correlation features, and map the state features to nodes of the topological structure and the correlation features to edges of the topological structure to construct a dynamic causal graph.

[0032] The report generation module is used to input dynamic causal graphs into a spatiotemporal convolutional neural network model. The feature aggregation layer uses the adjacency normalization operator to perform node feature extraction and spatial relationship fusion. The spatiotemporal coupling layer performs multi-scale feature enhancement through the residual cross-layer connection mechanism to form an anomaly impact score. The anomaly impact score is then classified and structured to output an anomaly monitoring report.

[0033] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the online anomaly monitoring method for a linear moving cut slurry sampler as described in the first aspect of the present invention.

[0034] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the online anomaly monitoring method for a linear moving cutter slurry sampler as described in the first aspect of the present invention.

[0035] The beneficial effects of this invention are as follows: By employing sliding window statistical analysis and entropy weighting to process the slurry sampler operating condition dataset, multi-source data can be more fully integrated, thereby accurately capturing the interactions and potential correlations between data from different sources. Utilizing dynamic causal graphs and spatiotemporal convolutional neural network models for deep feature mining and spatial relationship fusion effectively characterizes the complex causal relationships of the slurry sampler under complex operating conditions, significantly improving the ability to identify complex failure modes and further enhancing the reliability of anomaly monitoring. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart for an online anomaly monitoring method for a linear moving cutting slurry sampler.

[0038] Figure 2 This is a schematic diagram of an online anomaly monitoring system for a linear moving cutting slurry sampler.

[0039] Figure 3 A flowchart for generating spatiotemporal anomaly indicator vectors.

[0040] Figure 4 A flowchart for constructing a dynamic causal graph. Detailed Implementation

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0043] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0044] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides an online anomaly monitoring method for a linear moving cutter slurry sampler, comprising the following steps:

[0045] S1. Perform sliding window statistical analysis on the slurry sampler operating condition dataset to obtain a multidimensional feature matrix. Perform dynamic weighted aggregation on the multidimensional feature matrix to form a spatiotemporal analysis data package.

[0046] Specifically, the operations include the following:

[0047] S1.1 Collect the slurry sampler operating condition dataset, which includes mechanical movement and cutting parameters, slurry physical property parameters, and slurry sampler electrical parameters.

[0048] The mechanical moving cutting parameters include data on displacement, cutting speed, motion acceleration, and positioning deviation. Displacement data is collected using a displacement sensor, cutting speed data is collected using a servo encoder, motion acceleration data is collected using an accelerometer, and positioning deviation data is collected using a laser rangefinder.

[0049] The physical properties of the slurry include slurry concentration, density, particle size distribution, flow rate, and temperature. Slurry concentration is collected using an online concentration meter, density is collected using a mass flow meter, particle size distribution is collected using a laser particle size analyzer, flow rate is collected using an ultrasonic flow meter, and temperature is collected using a temperature transmitter.

[0050] The electrical parameters of the slurry sampler include voltage, current, power, frequency, and motor winding temperature. Voltage is collected using a voltage sensor, current is collected using a current transformer, power is collected using a power analyzer, frequency is collected using a frequency meter, and motor winding temperature is collected using a temperature sensor.

[0051] The slurry sampler operating condition dataset can not only comprehensively reflect the equipment's operating status and slurry physical properties, but also improve the accuracy and efficiency of mineral geological exploration services.

[0052] S1.2. Preprocessing of the slurry sampler operating condition dataset: Specifically, for mechanical movement and cutting parameters, moving mean filtering is applied for noise suppression, and linear interpolation is used to fill in missing values, improving the integrity of the mechanical movement and cutting parameters. Simultaneously, Z-score standardization is used for dimensional scaling to ensure uniformity of dimensions. For slurry physical property parameters, wavelet transform is used for multi-level decomposition and noise removal to improve the data purity and resolution of slurry physical property parameters. Simultaneously, a sliding window is used for time alignment to ensure the synchronization and consistency of slurry physical property parameters. For slurry sampler electrical parameters, low-pass filtering is used for high-frequency noise removal, and a sliding window is used for data smoothing to enhance the continuity of slurry sampler electrical parameters. Simultaneously, the NTP timestamp synchronization protocol is used for time synchronization to ensure the time consistency of slurry sampler electrical parameters.

[0053] S1.3. Fast Fourier Transform (FFT) is applied to perform bandgap energy separation on the preprocessed slurry sampler condition dataset to form a time-frequency composite parameter sequence. Specifically, the preprocessed dataset is filtered and smoothed, and an ADC converter is used for digital-to-analog conversion to obtain a digital signal. Simultaneously, a Hanning window is applied to the digital signal to generate a windowed time-domain signal. FFT is then applied to the windowed time-domain signal for bandgap energy separation. Further, frequency band decomposition and bandpass filtering are performed on the windowed time-domain signal, outputting a signal containing amplitude and phase. The complex spectrum of the information is obtained; the core frequency band (0 to 500 Hz) is extracted and divided into 8 sub-bands. Energy integration is performed on each sub-band to obtain the frequency band energy; then the frequency band energy is normalized to obtain a standardized energy distribution. The frequency domain features of the standardized energy distribution are extracted and multi-dimensional feature recombination is performed to obtain a frequency domain feature vector; sliding window sampling is performed on the frequency domain feature vector to obtain time-frequency composite parameters, and dynamic weight allocation and weighted aggregation are performed on the time-frequency composite parameters to output the time-frequency composite parameter sequence.

[0054] S1.4 Perform sliding window statistical analysis on the time-frequency composite parameter sequence to obtain a multidimensional feature matrix. In the specific operation, a fixed-length time window (e.g., 60 seconds) is used to slide the window across the time-frequency composite parameter sequence to remove incomplete abnormal windows and output the valid data window. Within the valid data window, a moving exponential average is performed to obtain the standardized range, and the standardized range is aggregated by a sliding window to obtain the basic statistical parameters. The basic statistical parameter features are scaled and reorganized in dimensions to generate multi-scale features. The multi-scale features are then horizontally stitched together and aligned in dimensions to form a multidimensional feature matrix.

[0055] S1.5. Perform weight allocation and vector reconstruction on the multidimensional feature matrix to generate a weighted spatiotemporal vector; perform dynamic weighted aggregation on the weighted spatiotemporal vector to output a spatiotemporal analysis data package. Specifically, in the weight allocation stage, perform linear transformation and weight allocation on the multidimensional feature matrix to generate a weighted feature matrix. At the same time, perform sparsification, dimensionality reduction, and orthogonalization on the weighted feature matrix to output a weighted optimized feature set. In the vector reconstruction stage, perform tensor decomposition on the weighted optimized feature set to extract spatial correlation features, and perform feature space projection and three-dimensional tensor reconstruction on the spatial correlation features to form a weighted spatiotemporal vector. In the weighted aggregation stage, perform feature concatenation and weighted aggregation on the weighted spatiotemporal vector volume to obtain primary spatiotemporal aggregated data. Simultaneously, perform temporal alignment and data smoothing on the primary spatiotemporal aggregated data to output a spatiotemporal analysis data package.

[0056] S2. Use the entropy weighting method to perform confidence weighting calculation on the spatiotemporal analysis data packet to obtain the anomaly probability value, and use the time decay compensation factor to perform time-series correction and spatial correlation mapping on the anomaly probability value to generate a spatiotemporal anomaly index vector.

[0057] Specifically, the operations include the following:

[0058] S2.1 Perform synergistic analysis on the spatiotemporal analysis data package and output a trend synergistic interaction matrix. Specifically, uniform sampling is performed on the spatiotemporal analysis data package to generate a random feature point set. The K-nearest neighbor algorithm is used to perform neighborhood relationship analysis on the random feature point set. Furthermore, distance measurement and similarity ranking are performed on the random feature point set to generate a neighborhood relationship matrix. Weight allocation and feature aggregation are performed on the neighborhood relationship matrix to form a standardized correlation matrix. The cross-correlation method is used to perform synergistic analysis on the standardized correlation matrix. Furthermore, the standardized correlation matrix is ​​symmetric to obtain an optimized correlation matrix. Eigenvalue decomposition is performed on the optimized correlation matrix to decompose it into eigenvalues ​​and eigenvectors. Principal component extraction is performed simultaneously: the first three principal component features of the eigenvalues ​​and eigenvectors are extracted, and vector normalization and dimension reorganization are performed to form synergistic feature vectors. Tensor expansion and matrix transformation are performed on the synergistic feature vectors to obtain the trend synergistic interaction matrix.

[0059] S2.2. The entropy weight method is used to perform risk quantification and contribution allocation on the trend collaborative interaction matrix, generating anomaly quantification parameters. These parameters are then weighted by confidence level to obtain anomaly probability values. Specifically, in the risk quantification stage, kernel density estimation is used to fit the probability density of the trend collaborative interaction matrix. Further, the trend collaborative interaction matrix is ​​discretized with equal width to obtain discretized feature intervals. Frequency statistics are performed on these discretized feature intervals to obtain feature frequency distributions. Simultaneously, the probability density of the feature frequency distribution is normalized to generate a probability distribution matrix. A logarithmic transformation is then performed on the probability distribution matrix to output a risk distribution vector. In the contribution allocation stage, the risk distribution vector is weighted and summed using matrix multiplication to obtain a weighted risk score. Based on the weighted risk score, the risk distribution vector is allocated a contribution value to form a risk contribution value vector. Linear combination and dimensionality fusion are then performed on the risk contribution value vector to obtain the anomaly quantification parameters.

[0060] In the weighted calculation stage, weights are assigned and parameters are aggregated for the anomaly quantification parameters to obtain a weighted parameter matrix. The entropy weight method is then applied to measure the information entropy of the weighted parameter matrix. Further, orthogonal decomposition and weighted averaging are performed on the weighted parameter matrix to obtain initial entropy weights. These initial entropy weights are then input into the Shannon entropy formula for confidence-weighted calculation to obtain the optimized information entropy value. The Softmax function is then applied to perform a nonlinear transformation on the optimized information entropy value, outputting the anomaly probability value. The specific mathematical formula is as follows.

[0061] ;

[0062] in, This represents the probability value of an anomaly. Represents the natural constant. This indicates the optimization of information entropy value. Represents the standard entropy value;

[0063] It should be noted that the standard entropy value is defined based on the median of the historical distribution of the optimized information entropy value.

[0064] S2.3. Utilize a time decay compensation factor to perform window smoothing and time weight allocation on the abnormal probability values, generating a decay correction sequence. Then, perform time-series correction on the abnormal probability values ​​based on the decay correction sequence, generating a time-series correction vector. In specific operations, during the window smoothing stage, time alignment and mean filtering are performed on the abnormal probability values ​​to ensure time-series continuity, outputting a smoothed probability sequence. A dual sliding window strategy is adopted to extract the long-term trend change features and short-term fluctuation features of the smoothed probability sequence: the outer window performs trend extraction and feature aggregation on the smoothed probability sequence to capture medium- and long-term trend change features, while the inner window performs fluctuation extraction and noise suppression on the smoothed probability sequence to generate short-term fluctuation features.

[0065] In the time weighting stage, the medium- and long-term trend change characteristics and short-term fluctuation characteristics are fused and weighted to obtain a weighted feature vector. The weighted feature vector is dynamically decayed and quantified using a time decay compensation factor to obtain the time correlation. When the time correlation is lower than the decay threshold, the weighted feature vector is determined to be recent data and needs to be assigned a higher time weight (e.g., 0.7). When the time correlation is higher than the decay threshold, the weighted feature vector is determined to be distant data and needs to be assigned a lower time weight (e.g., 0.3). According to the assigned time weight, the weighted feature vector is subjected to linear combination and dynamic adjustment to obtain a time-weighted sequence. The time-weighted sequence is then processed by a moving average to output a decay correction sequence.

[0066] It should be noted that the time decay compensation factor is defined based on the dynamic decay characteristics of the abnormal probability value, and its value range is [0.1, 0.9]; the decay threshold is defined based on the historical distribution characteristics of the weighted feature vector, and its value range is [0.4, 0.6].

[0067] Based on the decay correction sequence, time-series correction is performed on the anomaly probability value. Further, trend decomposition is performed on the decay correction sequence to obtain trend feature parameters. The trend feature parameters are dynamically weighted and numerically normalized to generate trend correction values. Based on the trend correction values, deviation compensation is performed on the anomaly probability value to obtain the correction probability value. The correction probability value is then subjected to smoothing filtering and nonlinear transformation to generate a time-series correction vector.

[0068] S2.4. Perform region partitioning and feature aggregation on the time-series correction vector to obtain a spatial feature set, and perform spatial association mapping on the spatial feature set to generate a spatiotemporal anomaly index vector. In specific operations, similarity measurement is performed on the time-series correction vector to form an initial association matrix. The initial association matrix is ​​then symmetricized to obtain an adjacency weight matrix. Feature propagation and weight aggregation are performed on the adjacency weight matrix to construct a spatial relationship graph. The spatial relationship graph is partitioned using spectral decomposition. Furthermore, eigenvalue decomposition and projection transformation are performed on the spatial relationship graph to obtain partition boundaries. Smoothing and weight adjustment are performed on the partition boundaries to generate optimized partitions. Finally, the optimized partitions are merged to obtain the initial spatial partitions.

[0069] The initial spatial partitions are subjected to feature aggregation. Further, the initial spatial partitions are decomposed into multi-scale features to obtain multi-level spatial features. Texture features are extracted and geometric features are described on the multi-level spatial features to generate regional feature vectors. Variance comparison and trend matching are performed on the regional feature vectors to obtain a set of regional features. At the same time, spatial interpolation and feature fusion are performed on the set of regional features to output a set of spatial features. Multi-dimensional feature projection is performed on the set of spatial features to obtain a latent spatial representation. Spatial correlation mapping is performed on the latent spatial representation to obtain a spatiotemporal anomaly index vector.

[0070] S3. Extract the state features and correlation features of the spatiotemporal anomaly index vector, map the state features to nodes of the topological structure, and map the correlation features to edges of the topological structure to construct a dynamic causal graph.

[0071] Specifically, the operations include the following:

[0072] S3.1. The correlation strength of the spatiotemporal anomaly index vector is analyzed to form a correlation strength matrix. Singular value decomposition is then performed on the correlation strength matrix to extract state and correlation features. Specifically, in the correlation strength analysis stage, the Pearson correlation coefficient is used to measure the covariance of the spatiotemporal anomaly index vector, generating a covariance matrix. The standard deviation of the covariance matrix is ​​then normalized to obtain the linear correlation coefficient. Next, polynomial fitting is performed on the linear correlation coefficient. Further, the linear correlation coefficient is discretized to obtain the basic fitting parameters. Weight allocation and B-spline interpolation are performed on the basic fitting parameters to obtain a smoothed intensity feature curve. The smoothed intensity feature curve is then normalized to generate the original correlation strength distribution. Simultaneously, dynamic weighting and dimension fusion are performed on the original correlation strength distribution to output an initial correlation strength vector. This initial correlation strength vector is then regularized and matrix transformed to eliminate noise interference and ensure stability, forming the correlation strength matrix.

[0073] It should be noted that the Pearson correlation coefficient is based on the definition of linear correlation of spatiotemporal anomaly index vectors, and its value range is [-1, 1].

[0074] In the singular value decomposition stage, the correlation strength matrix is ​​subjected to eigenvalue decomposition using the Lanczos algorithm. Further, the top k eigenvalues ​​of the correlation strength matrix are extracted and sorted by dimensionality reduction to obtain a salient feature set. This salient feature set is then orthogonally decomposed to generate left and right singular vectors. The left singular vector is rotated to enhance its orthogonality, resulting in an optimized state vector. Principal component projection is then performed on the optimized state vector to obtain state features. Next, the right singular vector undergoes spatial standardization to ensure scale consistency, outputting a standardized correlation vector. This standardized correlation vector is then weighted and aggregated to generate correlation features. The state features reflect the time-varying pattern of the overall operation mode of the slurry sampler, while the correlation features characterize the coupling strength between different spatial units of the slurry sampler.

[0075] S3.2. Normalize the state features to obtain standard state parameters, perform node attribute mapping on the standard state parameters, and generate nodes of the topology structure. In the specific operation, during the dimension normalization stage, feature alignment is performed on the state features, and MinMax normalization is used for dimension normalization. Simultaneously, linear transformation is used to map the feature values ​​of each dimension of the state features to the [0,1] interval to eliminate the difference in dimensions and retain the original distribution characteristics, thus obtaining the standard state parameters.

[0076] In the node attribute mapping stage, the equipment is spatially positioned based on the mechanical movement and cutting parameters. Further, the mechanical movement and cutting parameters are dynamically weighted to obtain weighted cutting parameters. Coordinate transformation and scale normalization are then performed on the weighted cutting parameters to obtain the equipment's spatial coordinates. The standard state parameters are then dimensionally aligned with the equipment's spatial coordinates, and features are stitched together to obtain a fused feature matrix. Principal component analysis is applied to the fused feature matrix to perform dimensional scaling and orthogonal rotation, obtaining node attribute parameters. These node attribute parameters are then spatially distributed and mapped using spatial interpolation to obtain the nodes of the topological structure.

[0077] S3.3. Quantize the interaction of associated features to generate an edge weight matrix. Map the edge weight matrix to obtain the edges of the topology. Specifically, perform interaction quantization on associated features. Further, perform distribution alignment and entropy fusion on associated features to generate an association coupling matrix. Perform feature decomposition on the association coupling matrix to obtain the interaction spectrum. Perform feature rotation and dimension compression on the interaction spectrum to generate an initial mutual information matrix. Perform dynamic weighting and scale normalization on the initial mutual information matrix to obtain the interaction intensity distribution. At the same time, perform sparsification on the interaction intensity distribution to obtain the edge weight matrix.

[0078] In the edge strength mapping stage, Euclidean distance measurement is performed on the spatial coordinates of the equipment to obtain the position difference matrix. Gaussian kernel transformation is then performed on the position difference matrix to obtain the actual physical distance of the equipment (such as pipe length and signal transmission loss). Logarithmic compression is then performed on the actual physical distance of the equipment to form a standard distance vector. The standardized distance vector and the edge weight matrix are then subjected to Hadamard integration to output a weighted correlation matrix. A distance attenuation factor is applied to the weighted correlation matrix to perform nonlinear scaling, thereby unifying the scale of the weight values ​​with the physical space. The edge strength parameters are then output and spatially propagated and mapped to generate the edges of the topology.

[0079] It should be noted that the distance decay factor is defined based on the spatial distribution characteristics of the weighted correlation matrix, and its value range is [0.5, 0.9].

[0080] S3.4. Perform dynamic hierarchical clustering on the nodes and edges of the topological structure to construct a dynamic causal graph. Specifically, density clustering is performed on the nodes to generate node clustering features; weight allocation and dimension aggregation are performed on the node clustering features to output the node clustering results; simultaneously, weight propagation and neighborhood aggregation are performed on the edges of the topological structure to generate edge association features, and nonlinear transformation and scale compression are performed on the edge association features to obtain a standardized causal matrix; dynamic hierarchical clustering is performed on the node clustering results and the standardized causal matrix; further, similarity fusion is performed on the node clustering results and the standardized causal matrix to output a hierarchical feature matrix; agglomerative clustering and pruning optimization are performed on the hierarchical feature matrix to generate an initial causal graph; topology simplification and noise filtering are performed on the initial causal graph to generate a dynamic causal graph.

[0081] S4. Input the dynamic causal graph into the spatiotemporal convolutional neural network model. The feature aggregation layer performs node feature extraction and spatial relationship fusion, the spatiotemporal coupling layer performs multi-scale feature enhancement, and outputs an anomaly monitoring report.

[0082] Specifically, the operations include the following:

[0083] S4.1 Construct and train a spatiotemporal convolutional neural network model. Specifically, in the PyTorch framework, the GraphConv network is called via the nn.Module parameter, and an adjacency normalization operator is embedded in the GraphConv network to perform symmetric normalization, ensuring stable propagation of node features. The input dimension of the GraphConv network is set to 64, the output dimension to 128, and the activation function to ReLU. BatchNorm1d is then applied after the GraphConv network for feature normalization to ensure stable gradient propagation. A Dropout layer is used for random deactivation to prevent overfitting, completing the feature aggregation layer. The 3D convolutional network is called via the nn.Conv3d function, and feature fusion is performed using the residual cross-layer connection mechanism to achieve stable transfer of deep features. The kernel size of the 3D convolutional network is set to 3×3×3, the stride to 1×1×1, and the padding mode to same. A LeakyReLU function is then applied after the 3D convolutional network for nonlinear transformation, completing the spatiotemporal coupling layer.

[0084] A multi-head attention mechanism is used to perform feature interaction between the feature aggregation layer and the spatiotemporal coupling layer to obtain cross-layer attention features. These cross-layer attention features are then concatenated to generate a fused feature matrix. A fully connected layer is used to perform dimensionality mapping on the fused feature matrix to obtain projected features. The projected features are then normalized using the Softmax function to generate attention weights. Based on these attention weights, the feature aggregation layer and the spatiotemporal coupling layer are stacked with cross-layer parameters, and feature enhancement is performed through residual connections to complete the construction of the spatiotemporal convolutional neural network model.

[0085] Next, the spatiotemporal convolutional neural network model is trained. Further, the node states and edge weights of the historical dynamic causal graph are extracted and integrated into time-series data, which is then divided into a sample set, a training set, and a validation set. On the sample set, data augmentation is used to apply random perturbations, and standardization is applied to unify the scale, forming augmented samples. On the training set, the Adam optimizer is used to update the parameters of the augmented samples, and an early stopping mechanism is simultaneously applied for training monitoring to obtain validation metrics. On the validation set, the validation metrics are averaged to obtain a smoothed validation curve. An error accumulation quantization is performed on the smoothed validation curve using a loss function to obtain the loss quantization value. When the loss quantization value exceeds the convergence threshold for five consecutive rounds, training terminates, and the trained spatiotemporal convolutional neural network model is output simultaneously.

[0086] It should be noted that the convergence threshold is defined based on the dynamic rate of change of the loss quantization value, and the value range is [0.001, 0.01].

[0087] S4.2 The feature aggregation layer applies the adjacency normalization operator to perform node feature extraction and spatial relationship fusion, generating a primary node embedding representation. Specifically, the dynamic causal graph is input into the spatiotemporal convolutional neural network model through a data loader. Feature alignment and weight allocation are performed on the dynamic causal graph, outputting initial node features. A nonlinear transformation is applied to the initial node features to obtain enhanced features. The adjacency normalization operator is applied to dynamically adjust and limit the range of the enhanced features to obtain scale-stable features. Matrix factorization is then performed on the scale-stable features, outputting a normalized adjacency matrix. An attention mechanism is used to propagate features and aggregate neighborhoods between the normalized adjacency matrix and the initial node features, outputting aggregated features. Spatial relationship fusion is performed on the aggregated features, and a fully connected layer is applied for linear transformation to obtain transformed features. The ReLU function is used to perform nonlinear activation on the transformed features, and a BatchNorm layer is used for standardization to eliminate feature scale differences, generating a primary node embedding representation.

[0088] It should be noted that the adjacency normalization operator is based on the definition of node degree distribution in dynamic causal graphs.

[0089] S4.3 The spatiotemporal coupling layer enhances multi-scale features through a residual cross-layer connection mechanism to obtain an enhanced spatiotemporal feature matrix. Specifically, a two-layer three-dimensional convolutional network is used to extract spatiotemporal features from the primary node embedding representation. Furthermore, the first-layer three-dimensional convolutional network performs initial convolution and feature transformation on the primary node embedding representation and applies LeakyReLU for nonlinear activation to generate initial spatiotemporal features. The second-layer three-dimensional convolutional network performs secondary convolution and feature fusion on the initial spatiotemporal features to obtain deep spatiotemporal features, and then performs scale normalization on the deep spatiotemporal features to obtain standard spatiotemporal features.

[0090] A residual cross-layer connection mechanism is employed to enhance standard spatiotemporal features at multiple scales. Further, channel attention is calculated on the standard spatiotemporal features, and the weights are normalized using the Sigmoid function to obtain an initial weight vector. This initial weight vector is then subjected to nonlinear transformation and feature compression using an MLP (Multilayer Perceptron) to generate channel weight values. The specific mathematical format is as follows.

[0091] ;

[0092] in, Indicates the channel weight value. Represents the normalization factor. The total number of channels representing standard spatiotemporal characteristics. Indicates the channel index. Indicates the first The initial weight vector of each channel;

[0093] It should be noted that the normalization factor is defined based on the output characteristics of the Sigmoid function, and its value range is (0,1).

[0094] Based on the channel weight values, feature weighting and cross-layer fusion are performed on the standard spatiotemporal features to generate weighted spatiotemporal features. Residual connections are then performed on the weighted spatiotemporal features to obtain enhanced spatiotemporal features. Stride convolution is performed on the enhanced spatiotemporal features to achieve hierarchical feature extraction. At the same time, dilated convolution is used to expand the receptive field and output multi-scale spatiotemporal features. LayerNorm is used to perform standardization and feature scaling on the multi-scale spatiotemporal features to output an enhanced spatiotemporal feature matrix.

[0095] S4.4. Perform fully connected transformation and probability normalization on the primary node embedding representation and the enhanced spatiotemporal feature matrix to form an anomaly impact score. Classify and structure the anomaly impact score to output an anomaly monitoring report. Specifically, in the fully connected transformation stage, the primary node embedding representation and the enhanced spatiotemporal feature matrix are concatenated to generate multidimensional fusion features. A fully connected transformation is then performed on the multidimensional fusion features to obtain transformed features, and nonlinear activation is applied to the transformed features to obtain a high-dimensional feature representation. In the probability normalization stage, a 1×1×1 convolution is used to randomly sample the high-dimensional feature representation to generate probability distribution parameters. Gaussian smoothing is applied to the probability distribution parameters to obtain normalized probabilities, and weighted aggregation is performed on the normalized probabilities to generate an anomaly impact score.

[0096] Anomaly impact scores are graded using multi-level thresholds. Further, multi-level thresholds are defined based on the statistical distribution of the anomaly impact scores. For example, the range of the first-level threshold is set to [0.8, 1.0]; the range of the second-level threshold is set to [0.6, 0.8]; and the range of the third-level threshold is set to [0.4, 0.6]. When the anomaly impact score falls within the range of the first-level threshold, it is defined as an urgent anomaly requiring immediate handling; when the anomaly impact score falls within the range of the second-level threshold, it is defined as a serious anomaly requiring priority handling; and when the anomaly impact score falls within the range of the third-level threshold, it is defined as a general anomaly requiring periodic handling. The graded levels are then prioritized and the graded anomaly results are output.

[0097] In the structured integration stage, feature alignment and weight allocation are performed on the hierarchical anomaly results to output an anomaly feature matrix; dimensionality compression and information fusion are performed on the anomaly feature matrix to obtain anomaly quantification parameters; the obtained anomaly quantification parameters are then structured and formatted to output an anomaly monitoring report.

[0098] This embodiment also provides an online anomaly monitoring system for a linear moving cutting slurry sampler, including:

[0099] The data analysis module is used to perform frequency band energy separation and sliding window statistical analysis on the slurry sampler operating condition dataset, obtain a multidimensional feature matrix, and perform weight allocation and dynamic weighted aggregation on the multidimensional feature matrix to form a spatiotemporal analysis data package;

[0100] The anomaly quantification module is used to perform collaborative analysis on spatiotemporal analysis data packets and output a trend collaborative interaction matrix. It uses the entropy weight method to perform risk quantification and contribution allocation on the trend collaborative interaction matrix, generates anomaly quantification parameters, and performs confidence weighted calculation on the anomaly quantification parameters to form anomaly probability values. It uses the time decay compensation factor to perform time-series correction and spatial correlation mapping on the anomaly probability values ​​to obtain a spatiotemporal anomaly index vector.

[0101] The graph construction module is used to perform correlation strength analysis and singular value decomposition on spatiotemporal anomaly index vectors, extract state features and correlation features, and map the state features to nodes of the topological structure and the correlation features to edges of the topological structure to construct a dynamic causal graph.

[0102] The report generation module is used to input dynamic causal graphs into a spatiotemporal convolutional neural network model. The feature aggregation layer uses the adjacency normalization operator to perform node feature extraction and spatial relationship fusion. The spatiotemporal coupling layer performs multi-scale feature enhancement through the residual cross-layer connection mechanism to form an anomaly impact score. The anomaly impact score is then classified and structured to output an anomaly monitoring report.

[0103] This embodiment also provides a computer device applicable to the online anomaly monitoring method of a linear moving cutter slurry sampler, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the online anomaly monitoring method of the linear moving cutter slurry sampler as proposed in the above embodiment.

[0104] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0105] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the online anomaly monitoring method for a linear moving cut slurry sampler as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0106] In summary, this invention utilizes sliding window statistical analysis and entropy weighting to process slurry sampler operating condition datasets, enabling more comprehensive fusion of multi-source data and thus accurately capturing the interactions and potential correlations between data from different sources. By employing dynamic causal graphs and spatiotemporal convolutional neural network models for deep feature mining and spatial relationship fusion, it effectively characterizes the complex causal relationships of slurry samplers under complex operating conditions, significantly improving the ability to identify complex failure modes and further enhancing the reliability of anomaly monitoring.

[0107] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for online anomaly monitoring of a linear moving cutting slurry sampler, characterized in that: include, Band energy separation and sliding window statistical analysis were performed on the slurry sampler operating condition dataset to obtain a multidimensional feature matrix. Weight allocation and dynamic weighted aggregation were then performed on the multidimensional feature matrix to form a spatiotemporal analysis data package. Perform synergy analysis on the spatiotemporal analysis data package and output a trend synergy interaction matrix; use the entropy weight method to perform risk quantification and contribution allocation on the trend synergy interaction matrix, generate anomaly quantification parameters, and perform confidence weighted calculation on the anomaly quantification parameters to form anomaly probability values; The time decay compensation factor is used to perform window smoothing and time weight allocation on the anomaly probability value to generate a decay correction sequence. The anomaly probability value is then corrected in time according to the decay correction sequence to generate a time correction vector. The time correction vector is then divided into regions and features are aggregated to obtain a spatial feature set. The spatial feature set is then spatially correlated and mapped to generate a spatiotemporal anomaly index vector. The spatiotemporal anomaly index vector is analyzed for correlation strength and decomposed by singular value, state features and correlation features are extracted, and the state features are mapped to nodes of the topological structure and the correlation features are mapped to edges of the topological structure to construct a dynamic causal graph. The dynamic causal graph is input into the spatiotemporal convolutional neural network model. The feature aggregation layer uses the adjacency normalization operator to perform node feature extraction and spatial relationship fusion. The spatiotemporal coupling layer performs multi-scale feature enhancement through the residual cross-layer connection mechanism to form an anomaly impact score. The anomaly impact score is then classified and structured to output an anomaly monitoring report.

2. The online anomaly monitoring method for a linear moving cutting slurry sampler as described in claim 1, characterized in that: The slurry sampler operating condition dataset includes mechanical movement and cutting parameters, slurry physical property parameters, and slurry sampler electrical parameters.

3. The online anomaly monitoring method for a linear moving cutting slurry sampler as described in claim 1, characterized in that: The process of forming the spatiotemporal analysis data packet specifically includes the following steps. Fast Fourier Transform is applied to perform frequency band energy separation on the slurry sampler operating condition dataset to form a time-frequency composite parameter sequence; sliding window statistical analysis is performed on the time-frequency composite parameter sequence to obtain a multidimensional feature matrix; Weighting and vector reconstruction are performed on the multidimensional feature matrix to generate a weighted spatiotemporal vector; dynamic weighted aggregation is performed on the weighted spatiotemporal vector to output a spatiotemporal analysis data package.

4. The online anomaly monitoring method for a linear moving cutting slurry sampler as described in claim 1, characterized in that: The construction of the dynamic causal graph specifically includes the following steps. The correlation strength of the spatiotemporal anomaly index vector is analyzed to form a correlation strength matrix. Singular value decomposition is then performed on the correlation strength matrix to extract state features and correlation features. The state features are normalized in dimension to obtain standard state parameters. Node attribute mapping is then performed on the standard state parameters to generate nodes in the topology structure. The interaction of associated features is quantified to generate an edge weight matrix. The edge strength of the edge weight matrix is ​​then mapped to obtain the edges of the topological structure. Dynamic hierarchical clustering is performed on the nodes and edges of the topology to construct a dynamic causal graph.

5. The online anomaly monitoring method for a linear moving cutting slurry sampler as described in claim 1, characterized in that: The spatiotemporal convolutional neural network model is constructed by building a feature aggregation layer and a spatiotemporal coupling layer, and applying a multi-head attention mechanism for cross-layer parameterization stacking.

6. The online anomaly monitoring method for a linear moving cutting slurry sampler as described in claim 5, characterized in that: The process of generating an abnormality impact score includes the following steps. The dynamic causal graph is input into the spatiotemporal convolutional neural network model. The feature aggregation layer applies the adjacency normalization operator to perform node feature extraction and spatial relationship fusion to generate a primary node embedding representation. The spatiotemporal coupling layer enhances multi-scale features through a residual cross-layer connection mechanism to obtain an enhanced spatiotemporal feature matrix; The primary node embedding representation and the enhanced spatiotemporal feature matrix are subjected to fully connected transformation and probability normalization to form an anomaly impact score.

7. An online anomaly monitoring system for a linear moving cutter slurry sampler, based on the online anomaly monitoring method for a linear moving cutter slurry sampler according to any one of claims 1 to 6, characterized in that: include, The data analysis module is used to perform frequency band energy separation and sliding window statistical analysis on the slurry sampler operating condition dataset, obtain a multidimensional feature matrix, and perform weight allocation and dynamic weighted aggregation on the multidimensional feature matrix to form a spatiotemporal analysis data package; The anomaly quantification module is used to perform collaborative analysis on spatiotemporal analysis data packets and output a trend collaborative interaction matrix. It uses the entropy weight method to perform risk quantification and contribution allocation on the trend collaborative interaction matrix, generates anomaly quantification parameters, and performs confidence weighted calculation on the anomaly quantification parameters to form anomaly probability values. The time decay compensation factor is used to perform window smoothing and time weight allocation on the anomaly probability value to generate a decay correction sequence. The anomaly probability value is then corrected in time according to the decay correction sequence to generate a time correction vector. The time correction vector is then divided into regions and features are aggregated to obtain a spatial feature set. The spatial feature set is then spatially correlated and mapped to generate a spatiotemporal anomaly index vector. The graph construction module is used to perform correlation strength analysis and singular value decomposition on spatiotemporal anomaly index vectors, extract state features and correlation features, and map the state features to nodes of the topological structure and the correlation features to edges of the topological structure to construct a dynamic causal graph. The report generation module is used to input dynamic causal graphs into a spatiotemporal convolutional neural network model. The feature aggregation layer uses the adjacency normalization operator to perform node feature extraction and spatial relationship fusion. The spatiotemporal coupling layer performs multi-scale feature enhancement through the residual cross-layer connection mechanism to form an anomaly impact score. The anomaly impact score is then classified and structured to output an anomaly monitoring report.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the online anomaly monitoring method for the linear moving cutting slurry sampler according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the online anomaly monitoring method for the linear moving cutting slurry sampler according to any one of claims 1 to 6.

Images