Multidimensional data acquisition and anomaly detection method for industrial equipment

Abstract

The present application relates to the field of data processing technology, and provides a multi-dimensional data acquisition and anomaly detection method for industrial equipment. By constructing a complete closed-loop system from multi-dimensional data acquisition, feature extraction, anomaly propagation probability judgment to combined judgment to generate anomaly diagnosis report, it breaks through the limitations of single indicators or fixed thresholds relied on by traditional industrial equipment detection. Through deep fusion of multi-source data and dynamic adaptive analysis, it significantly improves the accuracy and foresight of anomaly detection. A synergistic effect is formed between each step, which can adapt to the complex and changeable operating conditions of industrial equipment, effectively avoid false alarms and missed alarms, and provide a scientific basis for equipment maintenance.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a multi-dimensional data acquisition and anomaly detection method for industrial equipment. Background Art

[0002] In the field of intelligent operation and maintenance of industrial equipment, traditional anomaly detection methods primarily rely on single sensor data and fixed thresholds, or employ empirical rules to assess equipment status. While some solutions incorporate machine learning technology, these approaches often focus on analyzing surface features of data and struggle to cope with the complex and ever-changing operating conditions found in industrial scenarios. These approaches fail to address the problem of detecting and predicting anomaly propagation in industrial equipment through multi-dimensional data analysis combined with real-time operating conditions. Summary of the Invention

[0003] To address the shortcomings of the existing technology, the present application provides a method for multidimensional data acquisition and anomaly detection for industrial equipment. The method includes: performing modal analysis by combining the historical vibration spectrum of the industrial equipment with a three-dimensional model of the equipment, calculating the information entropy of each component in the industrial equipment, selecting detection point locations based on the information entropy of each component, deploying an Internet of Things gateway at the detection point location to obtain multidimensional data of the industrial equipment, and time-aligning the multidimensional data of the industrial equipment using a timestamp alignment algorithm to obtain a multidimensional data stream;

[0004] The vibration stiffness, thermal conductivity index, and current harmonics of the multidimensional data stream are extracted separately through a dual-channel network. A weight network is assigned according to the state of the industrial equipment. The vibration stiffness, thermal conductivity index, and current harmonics are weighted and fused together to generate a joint feature vector using the weight network.

[0005] Construct a normal state map of industrial equipment, use a clustering algorithm to determine the feature offset between the normal state map and the joint feature vector, and use a graph neural network to predict the abnormal diffusion path of industrial equipment and output an abnormal propagation probability map. The nodes of the graph neural network represent the components of the industrial equipment, and the edges of the graph neural network represent the abnormal propagation intensity.

[0006] The feature offset of the joint feature vector and the anomaly propagation probability map are combined for judgment, and the anomaly threshold is dynamically adjusted through Bayesian optimization. At the same time, an anomaly diagnosis report and anomaly confidence are output. A candidate detection point set is generated based on the anomaly propagation probability map, and the candidate detection point set is screened according to the anomaly confidence, and the detection point position is updated.

[0007] As an optional implementation, the logic for generating the joint feature vector includes:

[0008] Preprocess the multidimensional data stream and extract the vibration stiffness, thermal conductivity index and current harmonics of the multidimensional data stream through a dual-channel network;

[0009] The state of industrial equipment is comprehensively judged by the load, speed and temperature of the industrial equipment. A weight network is assigned based on the state of the industrial equipment through a multi-layer perceptron to output the weight coefficients of vibration stiffness, thermal conductivity index and current harmonics.

[0010] The vibration stiffness, thermal conductivity index and current harmonics are weightedly fused with the weight coefficients of the vibration stiffness, thermal conductivity index and current harmonics to generate a joint eigenvector.

[0011] As an optional implementation manner, the judgment logic of the feature deviation degree includes:

[0012] The joint feature vector is mapped to the normal state map through principal component analysis. The distance between the real-time operating parameters of the industrial equipment and the operating parameters in the normal state map is calculated through Euclidean distance. The operating condition feature cluster in the normal state map is selected as the matching cluster based on the distance.

[0013] In the matching cluster, the local neighborhood is determined with the mapped joint feature vector as the center and a hypersphere with a fixed radius as the range;

[0014] The Mahalanobis distance between each normal state feature point and the joint feature vector in the local neighborhood is calculated, and the average value of all Mahalanobis distances in the local neighborhood is used as the feature offset between the normal state map and the joint feature vector.

[0015] As an optional implementation, the output logic of the anomaly propagation probability map includes:

[0016] Build a graph neural network, where the nodes represent the components of industrial equipment and the edges represent the intensity of anomaly propagation. The feature offset and the status of each component are input into the graph neural network as node features.

[0017] Perform multiple graph convolution operations in the graph neural network, update the node features of each node, set an abnormality threshold, compare the node features of each node with the abnormality threshold to identify abnormal nodes and locate abnormal components, and connect abnormal nodes to obtain the abnormal diffusion path of industrial equipment;

[0018] For each abnormal node, the feature similarity with its neighboring nodes is determined by cosine similarity. The edge weights are corrected and normalized based on the abnormal propagation strength and feature similarity of the edges between nodes to obtain the abnormal propagation probability of each abnormal node to its neighboring nodes.

[0019] The anomaly propagation probability is used as the edge weight, the structure of the graph neural network is updated, and the output is visualized to obtain the anomaly propagation probability graph.

[0020] As an optional implementation, the adjustment logic of the abnormality threshold includes:

[0021] Extract the weight of each abnormal node and the edge connected to the abnormal node in the abnormal propagation probability graph, and merge it with the feature offset of the joint feature vector to obtain the comprehensive feature vector;

[0022] Calculate the prior probabilities of industrial equipment being in normal and abnormal states respectively, and statistically calculate the probability of the occurrence of comprehensive feature vectors in normal and abnormal states to obtain the likelihood function in normal state and the likelihood function in abnormal state;

[0023] According to Bayes' theorem, the posterior probability that the industrial equipment is in an abnormal state under the comprehensive feature vector is calculated, and the abnormal threshold is dynamically adjusted according to the posterior probability.

[0024] As an optional implementation manner, the detection point location selection logic includes:

[0025] By combining the historical vibration spectrum of industrial equipment with the three-dimensional model of the equipment for modal analysis, the vibration frequency band distribution of each component in the industrial equipment is obtained, and a vibration coupling matrix is ​​constructed. The matrix elements represent the transmission coefficient of vibration energy between components.

[0026] Based on the proportion of vibration energy of each component in different vibration frequency bands, the information entropy of each component is calculated, and the multi-scale entropy matrix is ​​obtained by determining the stability of the information entropy of each component at different time scales.

[0027] The detection points are selected according to the fluctuations in the multi-scale entropy matrix, and their locations are determined in combination with the three-dimensional model of the equipment.

[0028] As an optional implementation manner, the generation logic of the candidate detection point set includes:

[0029] Analyze the abnormal diffusion path in the abnormal propagation probability map and determine the range of abnormal components;

[0030] Conduct risk assessment on abnormal components based on their importance, abnormality level, and abnormality propagation probability;

[0031] According to the results of risk assessment, a risk threshold is set, the results of risk assessment are compared with the risk threshold, candidate detection points are screened out, and a candidate detection point set is generated.

[0032] As an optional implementation manner, the update logic of the detection point position includes:

[0033] Sort the components in the candidate detection point set in descending order based on the anomaly confidence;

[0034] Evaluate testing resources based on the number of IoT gateways and testing needs, and determine the number of testing points based on the testing resources;

[0035] According to the number of detection points, candidate detection points of the candidate detection point set are selected and the detection point positions are updated.

[0036] As an optional implementation, the output sub-logic of the abnormality confidence includes:

[0037] Calculate the correlation between the feature offset of the joint feature vector and the abnormal state, determine the contribution of the feature offset to the abnormality judgment, and obtain the feature offset contribution;

[0038] Analyze the contribution of abnormal propagation probability to abnormal judgment in the abnormal propagation probability graph to obtain the abnormal propagation probability contribution;

[0039] The feature offset contribution and the anomaly propagation probability contribution are weighted and fused to output the anomaly confidence of each component and generate an anomaly diagnosis report.

[0040] As an optional implementation, the dual-channel network includes a physical channel and a data channel, the physical channel is used to extract vibration stiffness, and the data channel extracts thermal conductivity index and current harmonics through a convolutional neural network.

[0041] Compared with the existing technology, the beneficial effects of this application are: by constructing a complete closed-loop system from multi-dimensional data acquisition, feature extraction, abnormal propagation probability judgment to combined judgment to generate abnormal diagnosis reports, it breaks through the limitations of single indicators or fixed thresholds that traditional industrial equipment detection relies on. Through deep fusion of multi-source data and dynamic adaptive analysis, it significantly improves the accuracy and foresight of abnormality detection, forms a synergistic effect between each step, can adapt to the complex and changeable operating conditions of industrial equipment, effectively avoids false alarms and missed alarms, and provides a scientific basis for equipment maintenance.

[0042] Vibration-based modal analysis and information entropy calculation can accurately locate key monitoring parts of equipment, avoid redundant points, reduce hardware deployment costs and data processing burdens, eliminate timing deviations in multi-source sensor data, ensure the synchronization of multi-dimensional data, and lay a solid foundation for subsequent feature analysis.

[0043] Vibration stiffness, thermal conductivity index, and current harmonics are extracted through a dual-channel network to achieve multi-dimensional and accurate characterization of the equipment's operating status. A weighted network is assigned based on the status of the industrial equipment, highlighting the contribution of key indicators to abnormality judgment and enhancing the detection system's adaptability to working conditions.

[0044] Build a normal state map of industrial equipment to avoid misjudgments caused by data congestion, provide an accurate comparison benchmark for anomaly detection, integrate equipment topology and feature offset through graph neural networks, visualize and predict anomaly diffusion paths, and help operation and maintenance personnel predict failure chain reactions in advance.

[0045] Dynamically adjust the anomaly threshold based on the posterior probability of anomalies, flexibly balance detection sensitivity and accuracy as working conditions change, reduce reliance on human intervention, screen high-risk components based on anomaly confidence and anomaly propagation probability, dynamically update the layout of detection points, achieve efficient configuration of detection resources, and form a virtuous cycle of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be derived from these drawings without inventive work. Among them:

[0047] Figure 1 A flowchart of a method for multi-dimensional data acquisition and anomaly detection for industrial equipment provided in an embodiment of the present application;

[0048] Figure 2 A logic diagram for determining the characteristic deviation of the multidimensional data acquisition and anomaly detection method for industrial equipment provided in an embodiment of the present application;

[0049] Figure 3 This is an output logic diagram of the anomaly propagation probability graph of the multi-dimensional data acquisition and anomaly detection method for industrial equipment provided in an embodiment of the present application;

[0050] Figure 4 This is a logic diagram for generating a candidate detection point set for the multidimensional data acquisition and anomaly detection method for industrial equipment provided in an embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described below in conjunction with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0052] like Figure 1 As shown, the embodiment of the present application provides a flowchart of a method for multi-dimensional data acquisition and anomaly detection for industrial equipment. The method for multi-dimensional data acquisition and anomaly detection for industrial equipment includes:

[0053] S1. Perform modal analysis by combining the historical vibration spectrum of industrial equipment with the three-dimensional model of the equipment, calculate the information entropy of each component in the industrial equipment, select the detection point location based on the information entropy of each component, deploy the IoT gateway at the detection point location to obtain the multi-dimensional data of the industrial equipment, and use the timestamp alignment algorithm to time-align the multi-dimensional data of the industrial equipment to obtain a multi-dimensional data stream.

[0054] Specifically, the selection logic of the detection point location includes:

[0055] By combining the historical vibration spectrum of industrial equipment with the three-dimensional model of the equipment for modal analysis, the vibration frequency band distribution of each component in the industrial equipment is obtained, and a vibration coupling matrix is ​​constructed. The matrix elements represent the transmission coefficient of vibration energy between components.

[0056] Based on the proportion of vibration energy of each component in different vibration frequency bands, the information entropy of each component is calculated, and the multi-scale entropy matrix is ​​obtained by determining the stability of the information entropy of each component at different time scales.

[0057] The detection points are selected according to the fluctuations in the multi-scale entropy matrix, and their locations are determined in combination with the three-dimensional model of the equipment.

[0058] Mining drills experience complex vibrations during operation, necessitating the location of key vibration sources for precise monitoring. Modal analysis, combined with historical vibration spectra and three-dimensional models, quantifies the vibration transmission relationship between components, providing a basis for inspection point selection. Three months of historical vibration spectrum data for the drill rig were obtained, with a sampling frequency of 10 kHz and covering the 0-5 kHz frequency band. Modal analysis was then performed on the drill rig's three-dimensional model, revealing the vibration frequency band distribution of 20 core components, including the drill rod, gearbox, and hydraulic pump. A 10×10 vibration coupling matrix was constructed. For example, the vibration energy transfer coefficient between the drill rod and gearbox was determined through a combination of experimental measurement and finite element simulation. The matrix elements range from 0.1 to 0.8, with higher values ​​indicating stronger coupling. This quantifies the vibration coupling relationship, preventing inspection points from missing high-risk areas and improving targeted monitoring.

[0059] The proportion of vibration energy in a single frequency band cannot reflect the vibration stability of a component, and multi-scale entropy is needed to capture the fluctuation patterns at different time scales. The component is divided into five frequency bands, such as 0-1kHz and 1-2kHz, and the energy proportion of each component in each frequency band is calculated. The information entropy is substituted into the existing information entropy formula and repeatedly calculated at time scales of 1s, 5s, and 10s to form a 5×3 multi-scale entropy matrix. For example, the entropy value of a gearbox fluctuates between 0.3-0.5 at a time scale of 1s and stabilizes at 0.35-0.45 at a time scale of 10s. This allows components with complex vibration patterns and poor stability to be identified, such as hydraulic pumps with high-frequency vibrations, to avoid misjudgment.

[0060] Based on entropy fluctuations, detection points that best reflect equipment anomalies are screened, and the feasibility of sensor installation is ensured by combining with three-dimensional models. The entropy fluctuation threshold is set, for example, the standard deviation of the fluctuation in the multi-scale entropy matrix is ​​greater than 0.1, thereby screening out five high-fluctuation components such as drill pipe connectors and hydraulic pump bearings. The three-dimensional model of the drilling rig is used to plan the sensor installation positions. For example, a three-axis acceleration sensor is arranged on the non-rotating surface of the drill pipe connector, and vibration and temperature sensors are installed near the hydraulic pump bearings. This reduces the redundancy of detection points and reduces deployment costs, while ensuring full coverage of key components, determining the deployment location of the IoT gateway, and providing a precise entry point for multi-dimensional data acquisition.

[0061] To obtain multi-source data such as vibration, temperature, and current in real time, the IoT gateway integrates multiple protocol-adaptive sensors to communicate with the cloud. Industrial-grade IoT gateways are deployed at five detection points, connecting vibration sensors, infrared temperature sensors, and current transformers, etc., to synchronously acquire multi-dimensional data at a frequency of 200Hz and store it in the temporary cache area of ​​the edge computing node. This enables unified access to multi-dimensional data, reduces data transmission latency, provides raw data for feature extraction, and time alignment ensures data timing consistency.

[0062] The differences in sampling frequencies of different sensors lead to data timing misalignment, requiring a unified time base. A hardware clock synchronization module based on the NTP protocol is used to add a timestamp to each data frame. A linear interpolation algorithm is implemented to resample the vibration and current data to a unified frequency. A multidimensional data stream is generated in timestamp order, thereby eliminating data timing deviation and improving the accuracy of subsequent feature analysis. The regularized multidimensional data stream is directly input into the dual-channel network for feature extraction.

[0063] S2. Vibration stiffness, thermal conductivity index, and current harmonics of the multidimensional data stream are extracted separately through a dual-channel network. A weight network is assigned according to the state of the industrial equipment. The vibration stiffness, thermal conductivity index, and current harmonics are weightedly fused together with the weight network to generate a joint feature vector.

[0064] The dual-channel network includes a physical channel and a data channel. The physical channel is used to extract vibration stiffness, and the data channel extracts thermal conductivity index and current harmonics through a convolutional neural network.

[0065] During the drilling process of a mining drill, the drill rod and the rock formation will have continuous impact, and the gearbox will vibrate when running at high speed. The change in its stiffness is the core indicator of the health status of the mechanical structure. Traditional vibration analysis is easily affected by environmental interference and requires a combination of physical models and signal processing to accurately extract the stiffness characteristics; three-axis acceleration sensors are installed at the key connections of the drill rod, including the connection between the drill rod and the drill bit and the drill rod coupling. At the same time, a one-way acceleration sensor is arranged at the gearbox bearing seat, and a magnetic base adsorption method is used to ensure a rigid connection with the equipment to avoid signal distortion caused by loose installation. When constructing the finite element model of the drilling rig dynamics, the nonlinear elastic modulus characteristics of the drill rod material are considered, and the frequency response function is performed in combination with the measured vibration signal. For example, the normal natural frequency of the drill pipe is 80Hz, which shifts to 75Hz when the stiffness drops by 10%. The local vibration stiffness is then inverted and calculated. For periodic interference sources such as the drilling rig engine and hydraulic pump, adaptive notch filtering technology is used, and the least mean square algorithm is used to dynamically adjust the notch center frequency and bandwidth. For example, when engine speed fluctuations are detected, the 150Hz notch frequency is corrected to the range of 148-152Hz in real time to ensure that the interference suppression depth is greater than 40dB. This makes it possible to quantify the stiffness changes of the vibration and capture the drill pipe fatigue trend three drilling cycles earlier than traditional Fourier analysis, avoiding drill pipe breakage accidents caused by the continuous decline in stiffness.

[0066] Long-term, high-pressure operation of the hydraulic system of a mining drill can easily lead to abnormal oil temperature. When the motor experiences a sudden load change, current harmonic distortion will occur. These hidden parameters require deep learning to identify early fault characteristics. An infrared thermal imager is deployed to scan the surface of the hydraulic oil tank, capturing a 640×480 resolution thermal image sequence. Using an inverse heat conduction algorithm, such as the finite difference method, combined with the thermal conductivity of the tank material and the ambient convection heat transfer coefficient, the internal oil temperature distribution of the tank is determined, and the thermal conductivity index is then calculated.

[0067] A Rogowski coil is installed at the motor input terminal to collect current signals. After conversion by a 16-bit ADC, the current signal is input into a 1D-CNN network. The network structure uses a stack of residual blocks and contains five convolutional layers and three maximum pooling layers. At the output layer, a fully connected layer is used to extract the amplitudes of the third, fifth, and seventh harmonics. The convolutional block attention module automatically enhances the feature weight of the current signal at the moment the hydraulic pump starts, randomly rotates and adjusts the brightness of the thermal image, and adds Gaussian white noise to the current signal. This enables the automatic extraction of the thermal conductivity index and current harmonics, thereby improving the fault detection rate.

[0068] Specifically, the generation logic of the joint feature vector includes:

[0069] Preprocess the multidimensional data stream and extract the vibration stiffness, thermal conductivity index and current harmonics of the multidimensional data stream through a dual-channel network;

[0070] The state of industrial equipment is comprehensively judged by the load, speed and temperature of the industrial equipment. A weight network is assigned based on the state of the industrial equipment through a multi-layer perceptron to output the weight coefficients of vibration stiffness, thermal conductivity index and current harmonics.

[0071] The vibration stiffness, thermal conductivity index and current harmonics are weightedly fused with the weight coefficients of the vibration stiffness, thermal conductivity index and current harmonics to generate a joint eigenvector.

[0072] The original data has significant dimensional differences and contains non-stationary noise, so it is necessary to unify the representation form and enhance the feature stability; the vibration stiffness is normalized to [0,1] through the maximum and minimum values, the thermal conductivity index and the amplitude of the current harmonics are Z-score standardized, the vibration signal is denoised by wavelet to remove high-frequency mechanical noise, the thermal image sequence uses median filtering to eliminate random noise from the infrared thermal imager, and the current signal is smoothed by a filter; thus, the weights of different modal data are balanced in the feature space, reducing the model training bias caused by dimensional differences and improving the subsequent fusion accuracy.

[0073] Under different operating conditions, such as hard rock drilling (high torque and low speed) and soft soil operations (low torque and high speed), the sensitivity of each feature to faults varies for mining drills, requiring dynamic weight adjustment to optimize detection performance. Data from the torque sensor, speed encoder, and hydraulic pressure sensor are acquired separately to construct a fuzzy rule base. For example, a torque greater than 30kNm and a speed less than 50rpm is considered a hard rock condition, while a torque less than 10kNm and a speed greater than 100rpm is considered a soft soil condition. The torque is converted into the load of the mining drill using existing technology. The MLP network is trained using the DRQN ​​reinforcement learning algorithm, with anomaly detection accuracy as the reward function. This yields a weighted network for different industrial equipment states, including weight coefficients for vibration stiffness, thermal conductivity index, and current harmonics. This improves the overall fault detection accuracy, especially reducing false alarms at the moment of switching between operating conditions.

[0074] Vibration stiffness, thermal conductivity index, and current harmonic features are integrated to form a unified device status representation, reducing data dimensionality while retaining key information. Vibration stiffness, thermal conductivity index, and current harmonics are weighted and fused with their corresponding weight coefficients to generate a joint feature vector. This reduces data dimensionality while ensuring information integrity and improves the real-time performance of anomaly detection. The joint feature vector is directly compared with the normal state map to accelerate feature offset calculation and anomaly diffusion prediction.

[0075] S3. Construct a normal state map of industrial equipment, determine the feature offset between the normal state map and the joint feature vector through a clustering algorithm, and predict the abnormal diffusion path of industrial equipment through a graph neural network. Output an abnormal propagation probability map. The nodes of the graph neural network represent the components of the industrial equipment, and the edges of the graph neural network represent the abnormal propagation intensity.

[0076] The operating status and characteristics of mining drills vary greatly under different working conditions. For example, when drilling in hard rock and soft soil, the load, speed, temperature and other parameters of the drill, as well as the vibration, heat conduction and current characteristics of each component are significantly different. In order to accurately determine whether the equipment is in a normal state, it is necessary to establish a normal state map for each working condition to provide an accurate comparison benchmark for real-time data.

[0077] During the long-term stable operation of industrial equipment, the generated joint feature vectors are continuously obtained and classified and stored according to different working conditions of industrial equipment (such as different loads, speeds and temperatures, etc.), such as the soft soil working condition and hard rock working condition mentioned above. For each working condition, the principal component analysis method is used to reduce the dimension of the joint feature vector under the working condition, and the main characteristic components that can represent the normal state of the industrial equipment under the working condition are extracted, such as the mechanical structure stability of the drilling rig, the thermal balance of the hydraulic system and the operating efficiency of the motor. These main characteristic components and the corresponding working condition parameters are integrated to construct the normal state map of the industrial equipment. Represent a series of feature clusters in a high-dimensional space. Each operating condition feature cluster represents the normal state of the equipment under a certain operating condition. After the normal state map is constructed, the normal state map is visualized so that operators can intuitively understand the normal state distribution under different operating conditions. The divided operating condition modeling avoids misjudgment caused by the mixing of data from different operating conditions, and can more accurately reflect the normal state of the equipment under different operating conditions. At the same time, the data dimension is compressed through dimensionality reduction processing, the calculation amount of subsequent feature comparison is reduced, the comparison efficiency is improved, and an accurate reference is provided for the calculation of feature offset, ensuring that the abnormal judgment can fit the actual working conditions, thereby improving the accuracy and reliability of anomaly detection.

[0078] Specifically, if Figure 2 As shown, the judgment logic of the feature deviation includes:

[0079] The joint feature vector is mapped to the normal state map through principal component analysis. The distance between the real-time operating parameters of the industrial equipment and the operating parameters in the normal state map is calculated through Euclidean distance. The operating condition feature cluster in the normal state map is selected as the matching cluster based on the distance.

[0080] In the matching cluster, the local neighborhood is determined with the mapped joint feature vector as the center and a hypersphere with a fixed radius as the range;

[0081] The Mahalanobis distance between each normal state feature point and the joint feature vector in the local neighborhood is calculated, and the average value of all Mahalanobis distances in the local neighborhood is used as the feature offset between the normal state map and the joint feature vector.

[0082] The real-time working conditions of mining drills will have a significant impact on their feature performance. If real-time data is directly compared with the normal state data of all working conditions without working condition matching, it will lead to misjudgment. Therefore, it is necessary to first match the normal state cluster of the corresponding working condition to eliminate the interference of working condition differences on abnormal judgment; the working condition parameters such as torque, speed and temperature of the drill rig are collected in real time, and the Euclidean distance between the real-time working condition parameters and the working condition parameters of the center of each working condition cluster in the normal state spectrum is calculated. The working condition feature cluster with the closest distance is selected as the matching cluster, and the current joint feature vector is converted to the feature space of the matching working condition through the PCA mapping matrix, which is consistent with the mapping matrix used when constructing the normal state spectrum. When mapping, it is necessary to pay attention to the accuracy and stability of matrix operations to avoid numerical errors; through working condition matching and feature mapping, it can be ensured that the feature comparison is performed under the same working conditions, eliminating the interference of working condition differences on abnormal judgment, improving the effectiveness and accuracy of feature comparison, determining the scope of local neighborhood search, narrowing the object of Mahalanobis distance calculation, and improving the efficiency and accuracy of subsequent calculations.

[0083] There is a certain degree of fluctuation in the data in the normal state map. Even under the same operating conditions, the operating status of the equipment will have slight changes. In order to reduce the impact of accidental deviations on anomaly judgment, it is necessary to define a reasonable local neighborhood range and focus on normal state feature points similar to the current feature vector; with the mapped joint feature vector as the center, a hypersphere neighborhood is defined within the matching cluster. The radius of the hypersphere is determined by multiplying the standard deviation of the historical data in each principal component direction by a coefficient. When determining the neighborhood range, it is necessary to comprehensively consider the data distribution and the sensitivity requirements of anomaly detection. Normal state feature points located within the hypersphere neighborhood are screened out in the matching cluster. To improve screening efficiency, the feature points in the matching cluster are organized and managed through the KD tree; thereby avoiding the interference of extreme values ​​and accidental deviations on anomaly judgment, focusing on the normal feature distribution under similar operating conditions, improving the stability and reliability of anomaly judgment, limiting the objects of Mahalanobis distance calculation, reducing the amount of calculation, and improving the accuracy of feature offset calculation.

[0084] There are correlations between the features in the joint eigenvector. For example, vibration stiffness and current harmonics vary in tandem with the device's operating state. The Euclidean distance does not consider the correlation between features, while the Mahalanobis distance comprehensively considers the coupling relationship between features and more accurately measures the difference between two eigenvectors. The Mahalanobis distance is calculated between each normal feature point in the local neighborhood and the current eigenvector. The Mahalanobis distance calculation requires the covariance matrix of the feature points in the local neighborhood. When calculating the covariance matrix, the sample size and distribution of the data must be considered to avoid matrix singularity. To improve computational efficiency, the covariance matrix is ​​processed using matrix decomposition technology. The average of all Mahalanobis distances in the local neighborhood is taken as the feature deviation. When calculating the average, all Mahalanobis distances are weighted averaged. The weight is determined by the similarity between the normal feature points in the local neighborhood and the current joint eigenvector. The larger the feature deviation, the further the current joint eigenvector deviates from its normal state. This comprehensively considers the correlation between features, avoids misjudgments caused by single-dimensional distance calculations, and more accurately reflects the difference between the current and normal states of the device.

[0085] Specifically, if Figure 3 As shown, the output logic of the anomaly propagation probability graph includes:

[0086] Build a graph neural network, where the nodes represent the components of industrial equipment and the edges represent the intensity of anomaly propagation. The feature offset and the status of each component are input into the graph neural network as node features.

[0087] Perform multiple graph convolution operations in the graph neural network, update the node features of each node, set an abnormality threshold, compare the node features of each node with the abnormality threshold to identify abnormal nodes and locate abnormal components, and connect abnormal nodes to obtain the abnormal diffusion path of industrial equipment;

[0088] For each abnormal node, the feature similarity with its neighboring nodes is determined by cosine similarity. The edge weights are corrected and normalized based on the abnormal propagation strength and feature similarity of the edges between nodes to obtain the abnormal propagation probability of each abnormal node to its neighboring nodes.

[0089] The anomaly propagation probability is used as the edge weight, the structure of the graph neural network is updated, and the output is visualized to obtain the anomaly propagation probability graph.

[0090] Complex physical connections exist between the various components of a mining drill rig. An anomaly in one component can trigger anomalies in other components. For example, a hydraulic pump failure can cause abnormal oil pipe pressure, which in turn affects the normal operation of the reversing valve. To accurately predict the anomaly propagation path and intensity, a graph neural network needs to be constructed to represent the associations between components. The graph neural network is constructed with the 20 core components of the mining drill rig (such as the drill rod, gearbox, and motor) as nodes of the graph neural network. The edge weights are initialized to the anomaly propagation intensity, which is represented by the vibration coupling intensity between components. This vibration coupling intensity is obtained from the vibration coupling matrix. When constructing the graph neural network, the node and edge attribute settings and the topological structure of the graph neural network need to be considered. Feature offset and the real-time status of the component (such as temperature and pressure sensor data) are input into the graph neural network as node features. To ensure data consistency and comparability, the node features are normalized and mapped to the [0, 1] interval. The graph neural network thus structurally expresses the topological structure of the equipment and the anomaly associations between components, providing an effective model foundation for anomaly propagation prediction.

[0091] During equipment operation, it is necessary to promptly and accurately locate abnormal nodes and identify abnormal diffusion paths so that appropriate maintenance measures can be taken to prevent further expansion of the abnormality. A three-layer graph convolution operation is performed in the graph neural network to update the node features of each node. The graph convolution operation updates the node features by aggregating the node's neighbor information, thereby capturing the association relationship between nodes. When performing the graph convolution operation, it is necessary to select an appropriate convolution kernel and activation function to improve the model's expressiveness. An abnormality threshold is set, such as a feature deviation greater than 0.8 is considered an abnormality. The node features of each node are compared with the abnormality threshold to identify abnormal nodes. When determining the abnormality threshold, it is necessary to comprehensively consider the equipment's historical operation data and the sensitivity requirements of anomaly detection. The abnormal nodes are connected through a depth-first search algorithm to form a preliminary abnormal diffusion path. During the search process, the access order and path information of each abnormal node are recorded. For example, the final abnormal diffusion path is "hydraulic pump → oil pipe → reversing valve". This can quickly locate the fault source and predict the direction of abnormal spread, providing an important basis for equipment maintenance decisions and helping to take timely measures to prevent the fault from expanding.

[0092] In order to quantify the possibility of anomalies propagating along the anomaly diffusion path and distinguish high-risk and low-risk components, it is necessary to calculate the anomaly propagation probability and generate an anomaly propagation probability graph so that operators can intuitively understand the risk distribution of anomaly propagation. For each anomaly node, calculate its cosine similarity with neighboring nodes to measure the feature similarity between nodes. Combined with the initial edge weight and the cosine similarity between nodes, the edge weight is corrected. The correction formula is: ,in represents the weight of the modified edge, Represents the weight coefficient, which is used to balance the impact of feature similarity and initial edge weight on the abnormal propagation intensity. Represents feature similarity, The initial edge weight is represented, and the corrected edge weight is normalized to obtain the anomaly propagation probability of each abnormal node to the neighboring node. The normalization process is performed through the softmax function to ensure that the sum of the propagation probabilities of all neighboring nodes is 1. Then, the structure of the graph neural network is updated, and the anomaly propagation probability is used as the edge weight. The graph neural network is visualized and output through visualization tools, and high-probability propagation paths are marked with bold lines; thus, the risk distribution of anomaly diffusion is intuitively presented, which helps operators formulate hierarchical maintenance strategies, give priority to inspecting and maintaining high-risk components, and improve the efficiency and pertinence of equipment maintenance. The anomaly propagation probability graph drives subsequent detection point updates and anomaly threshold adjustments, and provides important data support for intelligent decision-making.

[0093] S4. Combine the feature offset of the joint feature vector and the anomaly propagation probability map for judgment, and dynamically adjust the anomaly threshold through Bayesian optimization. At the same time, output the anomaly diagnosis report and anomaly confidence level, generate a candidate detection point set based on the anomaly propagation probability map, filter the candidate detection point set according to the anomaly confidence level, and update the detection point position.

[0094] Specifically, the adjustment logic of the abnormality threshold includes:

[0095] Extract the weight of each abnormal node and the edge connected to the abnormal node in the abnormal propagation probability graph, and merge it with the feature offset of the joint feature vector to obtain the comprehensive feature vector;

[0096] Calculate the prior probabilities of industrial equipment being in normal and abnormal states respectively, and statistically calculate the probability of the occurrence of comprehensive feature vectors in normal and abnormal states to obtain the likelihood function in normal state and the likelihood function in abnormal state;

[0097] According to Bayes' theorem, the posterior probability that the industrial equipment is in an abnormal state under the comprehensive feature vector is calculated, and the abnormal threshold is dynamically adjusted according to the posterior probability.

[0098] The working environment of mining drills is complex and changeable. Under different geological conditions, workloads, and equipment operating time, the characteristics of their normal and abnormal states can vary significantly. If a fixed anomaly threshold is used, false positives or missed positives are likely to occur when the operating conditions change. Therefore, it is necessary to combine real-time anomaly characteristics and historical data to dynamically adjust the anomaly threshold to adapt to the anomaly detection needs under different working conditions, thereby improving detection accuracy while ensuring detection sensitivity. From the anomaly propagation probability graph, the weights of each anomaly node and the edges connected to these anomaly nodes are extracted. For example, for the hydraulic pump anomaly node, its own anomaly propagation probability is 0.7, and the anomaly propagation probability of the edge connected to it is 0.6. At the same time, the characteristic offset of the joint feature vector is obtained, such as the offset of vibration stiffness is 0.4. This information is integrated to form a comprehensive feature vector containing 10 dimensions. During the integration process, different types of data need to be normalized to ensure that their weights in subsequent calculations are relatively reasonable. For example, the minimum and maximum normalization method is used to map all data to the interval [0,1].

[0099] Detailed operating data of mining drills for the past year was collected and analyzed in depth. The data were classified and counted according to normal and abnormal states. It was found through statistics that during the one-year operation time, the normal state accounted for about 90% of the time period and the abnormal state accounted for 10% of the time period. The Gaussian mixture model was used to fit the comprehensive feature vectors under normal and abnormal states. The Gaussian mixture model selected two components because it can better capture the complex distribution in the data. The likelihood function for the normal state obeys the Gaussian distribution, and the likelihood function for the abnormal state also obeys the Gaussian distribution. When constructing these distributions, the maximum likelihood estimation method is needed to determine the specific parameters of the corresponding mean vector and covariance matrix.

[0100] Based on the Bayesian theorem, the posterior probability of the device being in an abnormal state under the current comprehensive feature vector is calculated, and a set of segmented adjustment rules are designed to dynamically adjust the abnormal threshold according to the size of the posterior probability. For example, when the posterior probability of the device being in an abnormal state is less than or equal to 0.3, it means that the current device is less likely to be in an abnormal state. In order to reduce false alarms, the abnormal threshold is increased by 15%; when the posterior probability of the device being in an abnormal state is greater than or equal to 0.7, it means that the device is more likely to be in an abnormal state. In order to capture potential faults in time, the abnormal threshold is lowered by 20%; and when the posterior probability of the device being in an abnormal state is greater than 0.3 and less than 0.7, it means that the current state is not clear, and the abnormal threshold is kept. The normal threshold is stable, avoiding the uncertainty caused by frequent adjustments; this method of dynamically adjusting the abnormal threshold can adaptively match the real-time working conditions of the mining drill, avoiding the problem of detection failure due to sudden changes in working conditions. Under different working environments and load conditions, it can more accurately judge whether the equipment is in an abnormal state, thereby improving the reliability of abnormal detection. The dynamically adjusted abnormal threshold serves as an important benchmark for subsequent abnormal diagnosis, which will directly affect the calculation of abnormal confidence and the update strategy of the detection point. For example, when calculating the abnormal confidence, it is necessary to judge whether the feature deviation meets the abnormal standard based on the new abnormal threshold. When updating the detection point, it is also necessary to consider which components are more likely to have abnormalities under the new abnormal threshold.

[0101] Specifically, if Figure 4 As shown in Figure 2, the generation logic of the candidate detection point set includes:

[0102] Analyze the abnormal diffusion path in the abnormal propagation probability map and determine the range of abnormal components;

[0103] Conduct risk assessment on abnormal components based on their importance, abnormality level and abnormality propagation probability;

[0104] According to the results of risk assessment, a risk threshold is set, the results of risk assessment are compared with the risk threshold, candidate detection points are screened out, and a candidate detection point set is generated.

[0105] The components of a mining drill are closely interconnected. An anomaly in one component can spread to other components along a specific path. For example, a hydraulic pump failure can cause abnormal pressure in the oil pipe, which in turn affects the normal operation of the reversing valve. Therefore, it is necessary to analyze the anomaly propagation probability graph to identify the anomaly propagation path, identify high-risk areas and components, and rationally deploy detection resources in advance to avoid distributing detection resources on unnecessary components, thereby improving the efficiency and targetedness of detection. The anomaly propagation probability graph is deeply analyzed. By analyzing the relationship between nodes and edges in the anomaly propagation probability graph, strong correlation paths with high anomaly propagation probability are identified. For example, the total propagation probability of the path "hydraulic pump → oil pipe → reversing valve" is found to be greater than 0.8, indicating that the anomaly will spread along this anomaly diffusion path. During the analysis process, the shortest path algorithm in graph theory is used to determine the key propagation paths. Components are divided into different levels according to their role in the overall function of the mining drill rig. For example, the drill pipe is the core component directly involved in drilling and is classified as Level I, while the oil pipe is mainly responsible for transmitting hydraulic oil and is classified as Level II. Components of different levels have different weights in risk assessment.

[0106] The degree of component anomaly is quantified by combining the feature offset and the probability of anomaly propagation. The degree of anomaly is calculated by multiplying the two. At the same time, to more comprehensively assess component risk, the component's importance is combined with the degree of anomaly. For Level I components, their importance is higher, assuming it is 2. Then the result of its risk assessment is the degree of anomaly multiplied by the importance. A risk threshold is simultaneously set and the risk assessment results of all components are compared with this risk threshold. Components with risk assessment results greater than 0.2, such as oil pipes and reversing valves, are screened out and formed into a candidate detection point set. In this way, we can focus on high-risk components and concentrate detection resources on areas with the highest probability of anomalies, avoiding the dispersion and waste of detection resources and improving detection efficiency and accuracy. The generated candidate detection point set provides candidates for subsequent detection point updates. When updating the detection point positions, it is necessary to combine factors such as the anomaly confidence level to select appropriate components from the candidate detection point set to update the detection point positions to ensure continuous monitoring of high-risk components.

[0107] Specifically, the update logic of the detection point position includes:

[0108] Sort the components in the candidate detection point set in descending order based on the anomaly confidence;

[0109] Evaluate testing resources based on the number of IoT gateways and testing needs, and determine the number of testing points based on the testing resources;

[0110] According to the number of detection points, candidate detection points of the candidate detection point set are selected and the detection point positions are updated.

[0111] The development of mining drill failure is a dynamic process. In the early stage of the failure, only a certain component has a slight abnormality, but as time goes by, the abnormality will spread to other components. For example, in the early stage of the failure, only some minor problems occurred in the hydraulic pump, but if it is not handled in time, it will affect the normal operation of the motor in the later stage. Therefore, it is necessary to reallocate the detection point positions according to the real-time risk situation to ensure that the development and spread of the fault can be discovered in time, and to improve the pertinence and effectiveness of monitoring; combined with the abnormality confidence, the components concentrated in the candidate detection points are sorted in descending order. For example, if the abnormality confidence of the oil pipe is 0.7 and the abnormality confidence of the reversing valve is 0.6, then the oil pipe is placed in front. During the sorting process, efficient sorting algorithms such as quick sort are used to count the number of currently available IoT gateways. At the same time, according to the detection needs, it is determined that high-risk areas need to be fully covered and monitored. Taking these factors into consideration, the number of detection points for this update is determined to be 2. During the evaluation process, factors such as the IoT gateway's performance, detection range, and maintenance costs need to be considered. The top two components with the highest anomaly confidence levels—the oil pipeline and the reversing valve—are selected from the sorted candidate detection points. The IoT gateway is then redeployed to the locations of these components. During deployment, sensor installation parameters need to be updated simultaneously. For example, a pressure sensor can be added to the oil pipeline to more accurately monitor pressure changes within the pipeline. The newly deployed sensors also need to be calibrated and debugged to ensure proper operation. This method of dynamically updating detection point locations allows for timely response to fault evolution and ensures continuous monitoring of key nodes. Even if the fault situation changes, the detection system can always cover the components with the highest probability of anomalies, improving the timeliness and accuracy of monitoring.

[0112] Furthermore, the output sub-logic of the abnormal confidence includes:

[0113] Calculate the correlation between the feature offset of the joint feature vector and the abnormal state, determine the contribution of the feature offset to the abnormality judgment, and obtain the feature offset contribution;

[0114] Analyze the contribution of abnormal propagation probability to abnormal judgment in the abnormal propagation probability graph to obtain the abnormal propagation probability contribution;

[0115] The feature offset contribution and the anomaly propagation probability contribution are weighted and fused to output the anomaly confidence of each component and generate an anomaly diagnosis report.

[0116] In the process of abnormality detection of mining drills, it is necessary to quantitatively evaluate the abnormal possibility of each component in order to provide an intuitive and reliable basis for maintenance decisions; the Pearson correlation coefficient is used to calculate the degree of correlation between the characteristic deviation and the abnormal state. For example, through the analysis of historical data, the Pearson correlation coefficient between the vibration stiffness deviation and the abnormal event is calculated to be 0.6, which shows that the vibration stiffness deviation has a certain contribution to the abnormal judgment. The abnormal propagation probability of the abnormal node is directly taken as its contribution to the abnormal judgment. For example, the abnormal propagation probability of the oil pipeline is 0.6, then its abnormal propagation probability contribution is 0.6. In order to comprehensively consider the contribution of the characteristic deviation and the abnormal propagation probability contribution, the corresponding weights are set. The corresponding weights need to be obtained based on experiments. The abnormal confidence of each component is calculated by weighted summation. According to the calculated abnormal confidence, a detailed abnormal diagnosis report is generated. The report content includes a list of abnormal components, confidence ranking and recommended measures. The abnormal parts list refers to a list of all parts that are judged to be abnormal, such as oil pipes and reversing valves. The confidence ranking refers to ranking abnormal parts from high to low according to the abnormality confidence, which makes it convenient for maintenance personnel to quickly understand the abnormal possibility of each part. The recommended measures refer to giving corresponding recommended measures based on the abnormality confidence and the importance of the parts. For example, for parts with high abnormality confidence and importance level I, it is recommended to shut down for inspection first, while for parts with low abnormality confidence, it is recommended to strengthen monitoring. In this way, the abnormal risks can be presented in a structured manner, providing operation and maintenance personnel with clear and intuitive information to help them carry out graded disposal, thereby improving the scientific nature and efficiency of maintenance decisions. The abnormality confidence will be fed back to the detection point update and abnormality threshold adjustment. When the detection point is updated, the parts with high abnormality confidence will be given priority. When the abnormality threshold is adjusted, the distribution of abnormality confidence will also be referred to further optimize the abnormality threshold, thereby continuously optimizing the entire abnormality detection strategy.

Claims

1. A multidimensional data acquisition and anomaly detection method for industrial equipment, characterized in that: include: Modal analysis is performed by combining the historical vibration spectrum of industrial equipment with the three-dimensional model of the equipment. The information entropy of each component in the industrial equipment is calculated. The detection point location is selected based on the information entropy of each component. The IoT gateway is deployed at the detection point location to obtain multi-dimensional data of the industrial equipment. The multi-dimensional data of the industrial equipment is time-aligned using a timestamp alignment algorithm to obtain a multi-dimensional data stream. The selection logic of the detection point position includes: By combining the historical vibration spectrum of industrial equipment with the three-dimensional model of the equipment for modal analysis, the vibration frequency band distribution of each component in the industrial equipment is obtained, and a vibration coupling matrix is ​​constructed. The matrix elements represent the transmission coefficient of vibration energy between components. Based on the proportion of vibration energy of each component in different vibration frequency bands, the information entropy of each component is calculated, and the multi-scale entropy matrix is ​​obtained by determining the stability of the information entropy of each component at different time scales. Detection points are selected based on the fluctuations in the multi-scale entropy matrix, and their locations are determined in combination with the three-dimensional model of the equipment. The vibration stiffness, thermal conductivity index, and current harmonics of the multidimensional data stream are extracted separately through a dual-channel network. A weight network is assigned according to the state of the industrial equipment. The vibration stiffness, thermal conductivity index, and current harmonics are weighted and fused together to generate a joint feature vector using the weight network. Construct a normal state graph of industrial equipment, determine the feature deviation between the normal state graph and the joint feature vector using a clustering algorithm, and predict the abnormal diffusion path of industrial equipment using a graph neural network. Output an abnormal propagation probability graph. The nodes of the graph neural network represent the components of the industrial equipment, and the edges of the graph neural network represent the abnormal propagation intensity. The feature deviation judgment logic includes: The joint feature vector is mapped to the normal state map through principal component analysis. The distance between the real-time operating parameters of the industrial equipment and the operating parameters in the normal state map is calculated through Euclidean distance. The operating condition feature cluster in the normal state map is selected as the matching cluster based on the distance. In the matching cluster, the local neighborhood is determined with the mapped joint feature vector as the center and a hypersphere with a fixed radius as the range; Calculate the Mahalanobis distance between each normal state feature point and the joint feature vector in the local neighborhood, and take the average of all Mahalanobis distances in the local neighborhood as the feature offset between the normal state map and the joint feature vector; The feature offset of the joint feature vector and the anomaly propagation probability map are combined for judgment, and the anomaly threshold is dynamically adjusted through Bayesian optimization. At the same time, an anomaly diagnosis report and anomaly confidence are output. A candidate detection point set is generated based on the anomaly propagation probability map, and the candidate detection point set is screened according to the anomaly confidence, and the detection point position is updated.

2. The multidimensional data acquisition and anomaly detection method for industrial equipment according to claim 1, characterized in that: The generation logic of the joint feature vector includes: Preprocess the multidimensional data stream and extract the vibration stiffness, thermal conductivity index and current harmonics of the multidimensional data stream through a dual-channel network; The state of industrial equipment is comprehensively judged by the load, speed and temperature of the industrial equipment. A weight network is assigned based on the state of the industrial equipment through a multi-layer perceptron to output the weight coefficients of vibration stiffness, thermal conductivity index and current harmonics. The vibration stiffness, thermal conductivity index and current harmonics are weightedly fused with the weight coefficients of the vibration stiffness, thermal conductivity index and current harmonics to generate a joint eigenvector.

3. The multi-dimensional data acquisition and anomaly detection method for industrial equipment according to claim 2, characterized in that: The output logic of the anomaly propagation probability graph includes: Build a graph neural network, where the nodes represent the components of industrial equipment and the edges represent the intensity of anomaly propagation. The feature offset and the status of each component are input into the graph neural network as node features. Perform multiple graph convolution operations in the graph neural network, update the node features of each node, set an abnormality threshold, compare the node features of each node with the abnormality threshold to identify abnormal nodes and locate abnormal components, and connect abnormal nodes to obtain the abnormal diffusion path of industrial equipment; For each abnormal node, the feature similarity with its neighboring nodes is determined by cosine similarity. The edge weights are corrected and normalized based on the abnormal propagation strength and feature similarity of the edges between nodes to obtain the abnormal propagation probability of each abnormal node to its neighboring nodes. The anomaly propagation probability is used as the edge weight, the structure of the graph neural network is updated, and the output is visualized to obtain the anomaly propagation probability graph.

4. The multi-dimensional data acquisition and anomaly detection method for industrial equipment according to claim 3, characterized in that: The adjustment logic of the abnormal threshold includes: Extract the weight of each abnormal node and the edge connected to the abnormal node in the abnormal propagation probability graph, and merge it with the feature offset of the joint feature vector to obtain the comprehensive feature vector; Calculate the prior probabilities of industrial equipment being in normal and abnormal states respectively, and statistically calculate the probability of the occurrence of comprehensive feature vectors in normal and abnormal states to obtain the likelihood function in normal state and the likelihood function in abnormal state; According to Bayes' theorem, the posterior probability that the industrial equipment is in an abnormal state under the comprehensive feature vector is calculated, and the abnormal threshold is dynamically adjusted according to the posterior probability.

5. The multi-dimensional data acquisition and anomaly detection method for industrial equipment according to claim 4, characterized in that: The generation logic of the candidate detection point set includes: Analyze the abnormal diffusion path in the abnormal propagation probability map and determine the range of abnormal components; Conduct risk assessment on abnormal components based on their importance, abnormality level, and abnormality propagation probability; According to the results of risk assessment, a risk threshold is set, the results of risk assessment are compared with the risk threshold, candidate detection points are screened out, and a candidate detection point set is generated.

6. The multi-dimensional data acquisition and anomaly detection method for industrial equipment according to claim 5, characterized in that: The update logic of the detection point position includes: Sort the components in the candidate detection point set in descending order based on the anomaly confidence; Evaluate testing resources based on the number of IoT gateways and testing needs, and determine the number of testing points based on the testing resources; According to the number of detection points, candidate detection points of the candidate detection point set are selected and the detection point positions are updated.

7. The multi-dimensional data acquisition and anomaly detection method for industrial equipment according to claim 6, characterized in that: The output sub-logic of the abnormal confidence includes: Calculate the correlation between the feature offset of the joint feature vector and the abnormal state, determine the contribution of the feature offset to the abnormality judgment, and obtain the feature offset contribution; Analyze the contribution of abnormal propagation probability to abnormal judgment in the abnormal propagation probability graph to obtain the abnormal propagation probability contribution; The feature offset contribution and the anomaly propagation probability contribution are weighted and fused to output the anomaly confidence of each component and generate an anomaly diagnosis report.

8. The multi-dimensional data acquisition and anomaly detection method for industrial equipment according to claim 7, characterized in that: The dual-channel network includes a physical channel and a data channel. The physical channel is used to extract vibration stiffness, and the data channel extracts thermal conductivity index and current harmonics through a convolutional neural network.

Images