Digital monitoring method, system and device

Abstract

The application discloses a digital monitoring method, system and device, relates to the technical field of intelligent monitoring and digital twin intersection, and comprises the following steps: collecting multi-source data of equipment operation, wherein the multi-source data comprises one or more of video data, sensor data and operation and maintenance history data; inputting the multi-source data into a pre-trained multi-dimensional data correlation model for feature correlation analysis, and outputting fault type probability distribution and equipment residual life prediction results; wherein the multi-dimensional data correlation model is obtained through data preprocessing, feature extraction and correlation training on the multi-source data, can improve the accuracy, real-time performance and reliability of digital power plant equipment fault monitoring, and meets the intelligent operation and maintenance requirements of power plant equipment.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring and digital twin technology, and in particular to digital monitoring methods, systems and equipment. Background Technology

[0002] The current video surveillance fusion monitoring scheme integrates low frame rate 1080p video into a 3D scene through manual positioning to assist in observing the current status of equipment. The monitoring and maintenance of the equipment are performed by human judgment.

[0003] While the existing solutions mentioned above have achieved basic 3D visualization monitoring, they still have several technical shortcomings that make it difficult to meet the high-precision, high-real-time, and intelligent operation and maintenance requirements of digital power plants: First, the existing solutions only view equipment status data in a single way, lacking multi-source data correlation reasoning and fault root cause analysis, which makes them prone to false alarms and missed alarms, and cannot achieve proactive operation and maintenance; Second, when real-time video is integrated with digital twin models, the large amount of video texture data increases the pressure on video memory, and cross-system protocol compatibility is hindered.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a digital monitoring method, system, and equipment, which aims to solve the technical problem that existing technologies cannot meet the high-precision, high-real-time, and intelligent operation and maintenance requirements of digital power plants.

[0006] To achieve the above objectives, this application proposes a digital monitoring method, which includes: Collect multi-source data on the operation of the acquisition equipment, wherein the multi-source data includes one or more of video data, sensor data, and historical operation and maintenance data; The multi-source data is input into a pre-trained multi-dimensional data association model for feature association parsing, and the outputs the probability distribution of fault types and the prediction results of the remaining life of the equipment. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction and association training on the multi-source data.

[0007] In one embodiment, the step of inputting the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing and outputting the fault type probability distribution and equipment remaining life prediction results includes: Based on the video data, sensor data, and historical operation and maintenance data, the corresponding video feature vectors, sensor feature vectors, and knowledge graph feature vectors are extracted respectively. The video feature vector, sensor feature vector, and knowledge graph feature vector are concatenated into a multi-dimensional input feature vector, which is then input into a pre-trained multi-dimensional data association model. In the pre-trained multi-dimensional data association model, the association weight coefficients between each feature node in the multi-dimensional input feature vector are calculated through a multi-head graph attention mechanism; Based on the aforementioned correlation weight coefficients, each feature node is weighted and represented to enhance the expression of the correlation features between equipment operating parameters and equipment components, thereby obtaining weighted representation features. The weighted representation features are weighted, summed, and fused to generate a fused feature vector; Based on the fused feature vector, the probability distribution of fault types corresponding to various faults is calculated using an activation function, and the equipment fault status is determined by combining the preset probability threshold. The fused feature vector is input into a pre-trained Long Short-Term Memory (LSTM) network model for lifetime prediction mapping to obtain the device's remaining lifetime prediction result.

[0008] In one embodiment, before the step of inputting the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing and outputting the fault type probability distribution and equipment remaining life prediction results, the following steps are included: Frame sampling preprocessing and feature layer extraction are performed on the video data to obtain video feature vectors; Sensor data undergoes data preprocessing and time-frequency domain feature extraction to obtain sensor feature vectors; The historical operation and maintenance data is subjected to structured parsing, knowledge graph construction, and feature vectorization encoding to obtain knowledge graph feature vectors; The video feature vector, sensor feature vector, and knowledge graph feature vector are concatenated to construct a multi-dimensional training feature vector. The association relationship and association weight between each feature node in the multi-dimensional training feature vector are calculated through a multi-head graph attention mechanism to complete the weighted representation of the feature nodes and obtain the weighted representation features. The weighted representation features are fused from multiple heads, residual connections are introduced to optimize gradient propagation, and a random deactivation mechanism is used to suppress overfitting of the initial multidimensional data association model. The initial multidimensional data association model is iteratively optimized to obtain the trained multidimensional data association model.

[0009] In one embodiment, the step of performing frame sampling preprocessing and feature hierarchical extraction processing on video data to obtain video feature vectors includes: The video data is sampled frame by frame, and the sampled video data is preprocessed by denoising, contrast optimization, infrared temperature calibration and pixel normalization to obtain preprocessed video data. Based on a pre-trained residual convolutional neural network model, hierarchical feature extraction is performed on the preprocessed video data to obtain multi-level feature vectors. The multi-level feature vectors are compressed in dimension using global average pooling to obtain the initial video feature vectors. A channel attention mechanism is introduced to assign weights to the initial video feature vector, resulting in the final video feature vector.

[0010] In one embodiment, the step of performing data preprocessing and time-frequency domain feature extraction on the sensor data to obtain the sensor feature vector includes: The sensor data is subjected to outlier removal, missing value supplementation and filtering noise reduction to obtain preprocessed sensor data; Based on the preprocessed sensor data, time-domain feature parameters are extracted from the time-domain dimension, and frequency-domain feature parameters are extracted from the frequency-domain dimension through signal transformation. The time-domain feature parameters and the frequency-domain feature parameters are fused to generate a sensor feature vector.

[0011] In one embodiment, the step of performing structured parsing, knowledge graph construction, and feature vectorization encoding on historical operation and maintenance data to obtain knowledge graph feature vectors includes: The historical operation and maintenance data is subjected to structured parsing processing to identify and extract entities corresponding to faults, equipment components, and operation and maintenance measures, as well as the relationships between entities. Based on the relationships between entities, triplet data is generated. Based on the triplet data, the core entity types and attribute constraints are defined, and a knowledge graph is built by combining expert experience rules. The knowledge graph is vectorized by using a graph embedding algorithm to obtain the knowledge graph feature vector.

[0012] In one embodiment, after the step of vectorizing the knowledge graph using a graph embedding algorithm to obtain the knowledge graph feature vector, the method further includes: Get new operation and maintenance data; The incremental update algorithm is used to supplement the newly added operation and maintenance data into the constructed knowledge graph, and update the entity relationships and association weights in the knowledge graph.

[0013] In one embodiment, after the step of inputting the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing and outputting the fault type probability distribution and equipment remaining life prediction results, the method further includes: Based on the probability distribution of the fault types and the prediction results of the remaining lifespan of the equipment, the abnormal monitoring results corresponding to the video data, sensor data and operation and maintenance history data are obtained. The credibility weights of each anomaly monitoring result are weighted and fused to obtain the weighted score for each anomaly monitoring result. The weighted score is compared with the dynamic anomaly threshold, and the fault anomaly and false alarm samples are determined based on the comparison results. The dynamic anomaly threshold is obtained by statistical analysis of the equipment's historical normal operation data and adaptive adjustment of the equipment's real-time operating conditions. A false alarm case library is built based on the aforementioned fault anomalies and false alarm samples. The false alarm samples are labeled with false alarm causes, and statistical analysis is performed on the labeled false alarm samples to optimize the attention weight and dynamic anomaly threshold in the multi-dimensional data association model.

[0014] Furthermore, to achieve the above objectives, this application also proposes a digital monitoring system, wherein the digital monitoring includes: The acquisition module is used to collect multi-source data of equipment operation, including one or more of video data, sensor data, and historical operation and maintenance data. The inference module is used to input the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing, and output the probability distribution of fault types and the prediction results of the remaining life of the equipment. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction and association training on the multi-source data.

[0015] In addition, to achieve the above objectives, this application also proposes a digital monitoring device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the digital monitoring method as described above.

[0016] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the digital monitoring method described above.

[0017] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the digital monitoring method described above.

[0018] This application proposes a digital monitoring method, system, and device. The method includes: collecting multi-source data on equipment operation, wherein the multi-source data includes one or more of video data, sensor data, and historical maintenance data; inputting the multi-source data into a pre-trained multi-dimensional data association model for feature association analysis, and outputting a fault type probability distribution and a prediction result of the equipment's remaining lifespan. The multi-dimensional data association model is obtained by preprocessing, extracting features, and training the association on the multi-source data. This solution integrates video visual data, equipment sensor data, and historical maintenance data, and constructs a multi-dimensional data association model based on a graph attention neural network. It adaptively mines deep correlations among multi-source features, achieving accurate identification of equipment faults and time-series prediction of remaining lifespan. This effectively avoids the limitations of traditional monitoring methods with single data dimensions and fixed threshold judgment biases, significantly improving the accuracy, real-time performance, and reliability of fault monitoring in digital power plants, and meeting the intelligent operation and maintenance needs of power plant equipment. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating an embodiment of the digital monitoring method of this application. Figure 2 This is a flowchart illustrating Embodiment 2 of the digital monitoring method of this application; Figure 3 This is a flowchart illustrating Embodiment 3 of the digital monitoring method of this application; Figure 4 This is a flowchart illustrating Embodiment 4 of the digital monitoring method of this application; Figure 5 This is a schematic diagram of the module structure of the digital monitoring system according to an embodiment of this application; Figure 6 This is a schematic diagram of the device structure of the hardware operating environment involved in the digital monitoring method in this application embodiment.

[0022] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0023] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0024] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0025] The main solution of this application embodiment is: to collect multi-source data of equipment operation, including one or more of video data, sensor data and operation and maintenance history data; to input the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing, and to output the probability distribution of fault types and the prediction results of the remaining life of the equipment. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction and association training on the multi-source data.

[0026] In this embodiment, for ease of description, the following description uses a digital monitoring system as the executing entity.

[0027] While existing technologies have achieved basic 3D visualization inspection, they still have several technical shortcomings, making it difficult to meet the high-precision, high-real-time, and intelligent operation and maintenance requirements of digital power plants. 1. Weak AI intelligent analysis capabilities: It only views the device status data and lacks multi-source data correlation reasoning and fault root cause analysis, which makes it prone to false alarms and missed alarms and cannot achieve proactive operation and maintenance.

[0028] 2. Insufficient system adaptability and stability: When real-time video is fused with digital twin models, the large amount of video texture data increases the pressure on the video memory, and cross-system protocol compatibility is hindered.

[0029] This application provides a solution that constructs a multi-dimensional data association model of "video features + sensor data + operation and maintenance history", integrates deep learning algorithms to realize fault root cause analysis, false alarm filtering and equipment remaining life prediction, and breaks through the technical bottlenecks of traditional monitoring solutions such as single data dimension, weak intelligent analysis capability, poor system adaptability and low monitoring accuracy. It effectively improves the accuracy, real-time performance and system stability of digital power plant equipment monitoring, and adapts to the intelligent, proactive and refined operation and maintenance needs of power plants.

[0030] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions. The following description uses a personal computer as an example to illustrate this embodiment and the subsequent embodiments.

[0031] Based on this, the embodiments of this application provide a digital monitoring method, referring to... Figure 1 , Figure 1This is a flowchart illustrating the first embodiment of the digital monitoring method of this application.

[0032] In this embodiment, the digital monitoring method includes steps S10 to S40: Step S10: Collect multi-source data of device operation, including one or more of video data, sensor data, and historical operation and maintenance data; The video data is visual data collected in real time by the device, including visual information that can intuitively reflect the appearance and operating status of the device, such as the surface texture of the device, the movement trajectory of the components, and the infrared temperature distribution. The sensor data refers to the real-time operational parameter data collected by the sensing terminals deployed at various key points of the equipment, covering core operational characteristic parameters such as vibration, temperature, pressure, and rotational speed. This type of data has the characteristics of high-frequency acquisition, dynamic fluctuation, and susceptibility to interference from the environment and equipment operating conditions. The historical maintenance data refers to the maintenance archive data for the entire life cycle of the equipment, including but not limited to historical fault records, equipment repair work orders, component replacement ledgers, equipment cumulative runtime and operating condition calibration records, and other prior maintenance knowledge data.

[0033] Step S20: Input the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing, and output the fault type probability distribution and equipment remaining life prediction results. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction and association training on the multi-source data.

[0034] Understandably, existing solutions suffer from weak AI intelligent analysis capabilities, only able to view equipment status data individually, lacking the ability to perform deep correlation reasoning and fault root cause analysis of multi-source data. This easily leads to false alarms and missed alarms, making it impossible to achieve proactive predictive maintenance of equipment. Therefore, executing step S20 avoids the fault discrimination bias caused by the single dimension of traditional equipment monitoring data and insufficient feature correlation mining. This enables deep fusion and intelligent correlation analysis of multi-source heterogeneous data, accurately completing equipment fault identification and remaining life prediction, and supporting proactive intelligent maintenance of digital power plants.

[0035] In one feasible embodiment, step S20 may include steps S21 to S27: Step S21: Based on the video data, sensor data, and historical operation and maintenance data, extract the corresponding video feature vector, sensor feature vector, and knowledge graph feature vector respectively. In this embodiment, by using preset data preprocessing and feature extraction algorithms, hierarchical feature extraction and vectorization processing are performed on three types of multi-source data to obtain video feature vectors, sensor feature vectors and knowledge graph feature vectors with uniform dimensions and high feature recognition.

[0036] For sensor data, a 64-dimensional high-stability sensor feature vector is obtained through outlier removal, segmented filtering and noise reduction, and time-frequency domain feature extraction. For historical operation and maintenance data, a 64-dimensional standardized knowledge graph feature vector is obtained through structured parsing, triplet construction, knowledge graph construction, and TransE() graph embedding algorithm encoding, providing standardized input data for subsequent multi-feature fusion and correlation analysis.

[0037] Step S22: The video feature vector, sensor feature vector, and knowledge graph feature vector are concatenated into a multi-dimensional input feature vector and input into a pre-trained multi-dimensional data association model; In this embodiment, the 256-dimensional video feature vector, the 64-dimensional sensor feature vector, and the 64-dimensional knowledge graph feature vector are concatenated to construct a 384-dimensional multidimensional input feature vector. Each multidimensional input feature vector corresponds to a set of equipment operating status samples. The sample labels include the normal operating status and various fault operating statuses of the equipment, which fully represents the multidimensional working condition information of the equipment's visual status, operating parameters, and maintenance knowledge. The constructed multidimensional input feature vector is then input into a pre-trained multidimensional data association model.

[0038] The multi-dimensional data association model is a four-layer hierarchical architecture based on the graph attention neural network (GAT). From top to bottom, it consists of an input layer, a graph attention layer, a feature fusion layer, and an output layer. The parameters of each network layer are fixed and converged through batch sample iterative training, which enables multi-source feature association mining and intelligent parsing.

[0039] Step S23: In the pre-trained multi-dimensional data association model, the association weight coefficients between each feature node in the multi-dimensional input feature vector are calculated through a multi-head graph attention mechanism. In this embodiment, the graph attention layer in the multi-dimensional data association model has two attention heads, each configured with a single hidden layer structure and 128 hidden layer neurons. The association weight coefficients of each feature node within each multi-dimensional input feature vector are calculated using a self-attention mechanism. The specific calculation formula is as follows: α_ij=softmax_j(LeakyReLU(a^T[W h_i || W h_j])) Where w is a weight matrix of dimension 128×384, a is an attention coefficient vector of dimension 256×1, and || denotes vector concatenation. Based on this calculation method, the implicit strong correlation features between vibration peak values ​​and bearing components, infrared temperature and end cap components, and other equipment parameters and components are mined and identified, providing quantitative coefficient support for subsequent differentiated weighted characterization.

[0040] Step S24: Based on the association weight coefficient, perform weighted representation on each feature node to strengthen the expression of the association features between equipment operating parameters and equipment components, and obtain weighted representation features; In this embodiment, the weight coefficients of each feature node output by the graph attention layer of the multi-dimensional data association model are used to perform weighted representation processing on all feature nodes. The strong correlation features of equipment faults and core features of operating conditions are weighted and enhanced, while invalid features and interference noise features are weighted and suppressed. This strengthens the expression of deep correlation features between equipment operating parameters, key components and fault types, highlights the sensitive features of equipment faults, and finally completes the differentiated weighted representation of multi-dimensional feature nodes to obtain highly recognizable weighted representation features.

[0041] Step S25: Perform weighted summation and fusion on the weighted representation features to generate a fused feature vector; In this embodiment, the feature fusion layer of the multi-dimensional data association model performs weighted summation and fusion of multiple sets of weighted representation features output by the dual attention head, and completes deep feature aggregation based on the network structure solidified by offline training of the model, outputting a 128-dimensional fused feature vector.

[0042] Step S26: Based on the fused feature vector, the probability distribution of fault types corresponding to various faults is calculated using an activation function, and the equipment fault status is determined by combining the preset probability threshold. In this embodiment, the softmax activation function is used to output the probability distribution of fault types (such as bearing wear probability 0.92, gear crack probability 0.05, and normal state probability 0.03), and the probability of each fault type is compared with the preset probability threshold of 0.7. When the probability of a certain type of fault exceeds the threshold, it is determined that there is a corresponding fault.

[0043] Step S27: Input the fused feature vector into the pre-trained Long Short-Term Memory (LSTM) network model for lifetime prediction mapping to obtain the device remaining lifetime prediction result.

[0044] Simultaneously, the fused feature vectors are input into a pre-trained convergent long short-term memory network model. By utilizing the temporal evolution extrapolation capability of the long short-term memory network and combining the lifespan mapping relationship learned from the equipment's historical operation and fault samples, a temporal analysis of the equipment's operating trend is performed, and the equipment's remaining lifespan prediction results are output.

[0045] Through the above steps, a pre-trained multi-dimensional data association model is used to complete multi-source feature encoding, adaptive association weighting, feature fusion, and intelligent inference. Under the fixed parameter inference mode of the model, it can stably realize the output of probability distribution of multiple types of equipment faults, fault status determination, and time series prediction of remaining life. This solves the problems of single data dimension, lack of association analysis, and weak intelligent judgment capability of traditional monitoring methods, and effectively improves the accuracy and stability of fault monitoring and life prediction of digital power plant equipment.

[0046] The above-described embodiments collect multi-source data on equipment operation, including one or more of video data, sensor data, and historical maintenance data. This multi-source data is then input into a pre-trained multi-dimensional data association model for feature association analysis, outputting a fault type probability distribution and a predicted remaining lifespan of the equipment. The multi-dimensional data association model is obtained through data preprocessing, feature extraction, and association training of the multi-source data. This solution integrates video visual data, equipment sensor data, and historical maintenance data, and constructs a multi-dimensional data association model based on a graph attention neural network. It adaptively mines deep correlations among multi-source features, achieving accurate equipment fault identification and time-series prediction of remaining lifespan. This effectively avoids the limitations of traditional monitoring methods with single data dimensions and fixed threshold judgment biases, significantly improving the accuracy, real-time performance, and reliability of fault monitoring in digital power plants, and meeting the intelligent operation and maintenance needs of power plant equipment.

[0047] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 Before step S20, the digital monitoring method further includes steps S201 to S205: Step S201: Perform frame sampling preprocessing and feature layer extraction on the video data to obtain the video feature vector; Optionally, step S201 may also include steps S2011 to S2014: Step S2011: Frame sampling is performed on the video data, and the sampled video data is preprocessed by denoising, contrast optimization, infrared temperature calibration and pixel normalization to obtain preprocessed video data. In this embodiment, the original video data is first sampled frame by frame using a sampling strategy of 2-3 frames per second. This reduces the amount of data processing while preserving key operational information of the device, balancing algorithm efficiency with the integrity of the image information. Then, the sampled frames are denoised using a Gaussian filtering algorithm with a convolution kernel size of 3×3 and a standard deviation σ=0.8, effectively removing noise and key interference points from the video image. A histogram equalization algorithm is used to optimize the image grayscale contrast and enhance the detailed features of key areas of the device. For infrared monitoring video, temperature calibration is performed based on a blackbody radiation source to control the infrared temperature measurement error within ±0.5℃, unifying the mapping relationship between image pixels and actual temperature. Simultaneously, the image is normalized by using the formula x_norm=(x-x_min) / (x_max-x_min) (where x is the original gray value of the current pixel, x_min is the minimum gray value of a single frame, and x_max is the maximum gray value of a single frame). This normalizes the image pixel values ​​to the range of [0,1], eliminating image data deviations caused by changes in external lighting and differences in device shooting resolution, thus obtaining preprocessed video data.

[0048] Step S2012: Based on the pre-trained residual convolutional neural network model, perform hierarchical feature extraction on the preprocessed video data to obtain multi-level feature vectors; In this embodiment, a pre-trained ResNet-18 residual convolutional neural network model is used to extract shallow and deep features from the pre-processed video data. For shallow feature extraction, the first three convolutional blocks of the ResNet-18 model (Conv2d, BatchNorm, and ReLU) are used to extract low-dimensional features, including the edges of the device surface texture, component outlines, and infrared temperature gradient distribution. For deep feature extraction, the residual block structure is used to overcome the gradient vanishing problem and extract high-dimensional semantic features, such as the morphology of abnormal heating areas on bearing end caps, changes in gear meshing clearance, and image blurring features caused by component vibration. The final output is a 256-dimensional multi-level feature vector.

[0049] Step S2013: The multi-level feature vectors are compressed in dimension by global average pooling to obtain the initial video feature vectors; In this embodiment, global average pooling (GAP) is used to compress the global dimension and filter features of the multi-level feature vectors. The global information of each channel of the feature map is aggregated by mean statistics, and redundant and invalid local fragmented features are discarded. Core effective feature information such as device appearance defects and temperature anomalies is retained to the greatest extent, reducing the number of feature parameters and computational overhead, realizing lightweight compression of feature dimensions, and outputting an initial video feature vector with regular dimensions.

[0050] Step S2014: Introduce a channel attention mechanism to assign weights to the initial video feature vector to obtain the final video feature vector.

[0051] In this embodiment, the SE-Net channel attention mechanism is introduced to adaptively optimize the feature weights of the initial video feature vector. Through channel-dimensional squeezing and excitation operations, the importance of each feature channel is automatically learned, and the channel weight ratio of key fault features such as abnormal temperature, surface cracks of equipment, and component deformation is specifically strengthened. At the same time, the weight response of invalid and irrelevant features such as complex background interference, normal wear and tear of equipment, and environmental noise is suppressed, which significantly improves the distinguishability between effective fault features and interference features, and outputs the final video feature vector.

[0052] Step S202: Perform data preprocessing and time-frequency domain feature extraction on the sensor data to obtain the sensor feature vector; Optionally, step S202 may also include steps S2021 to S2023: Step S2021: Perform outlier removal, missing value supplementation, and filtering and noise reduction on the sensor data to obtain preprocessed sensor data. In this embodiment, outliers are first removed from the sensor data using the 3σ criterion; that is, sensor data is considered abnormal when it deviates from the mean by three times the standard deviation. Then, for missing values ​​caused by sensor disconnection, a combination of linear interpolation (missing duration < 5s) and LSTM (Long Short-Term Memory) prediction interpolation (missing duration ≥ 5s) is used to supplement the missing values, ensuring data continuity. For vibration signals in the sensor data, wavelet filtering algorithms (such as db4 wavelet basis, 5 decomposition layers) are used to separate high-frequency noise from low-frequency effective signals, preserving the impact characteristics and harmonic components of the vibration signal. For slowly varying sensor data with parameters such as temperature and pressure, a Kalman filter algorithm is used for dynamic noise reduction and error correction. The Kalman filter state equation is: X(k) = AX(k-1) + BU(k) + W(k) The observation equation is: Z(k) = HX(k) + V(k) Where X(k) is the system state variable at time k, A is the state transition matrix, B is the control input matrix, U(k) is the control input at time k, W(k) is the system process noise, Z(k) is the observation value at time k, H is the observation matrix, and V(k) is the observation noise.

[0053] By iteratively calculating the optimal state estimate through Kalman filtering, random measurement errors and environmental interference errors in the sensing and acquisition process are offset, effectively improving the smoothness and stability of the slowly changing operating parameters, and finally obtaining high-precision, missing, and low-noise preprocessed sensor data.

[0054] Step S2022: Based on the preprocessed sensor data, extract time-domain feature parameters from the time domain dimension and extract frequency-domain feature parameters from the frequency domain dimension through signal transformation. In this embodiment, feature parameters are extracted from both the time and frequency domains. The time-domain feature parameters include: vibration peak value, peak-to-peak value, mean, variance, kurtosis, and skewness (reflecting the signal distribution characteristics; kurtosis > 3 indicates the presence of an impact fault). The frequency-domain feature parameters are obtained by converting the time-domain signal into a frequency-domain signal using Fast Fourier Transform (FFT) to extract the fundamental frequency, harmonic amplitude, and frequency band energy proportion (such as the energy proportion corresponding to the bearing fault characteristic frequency).

[0055] Step S2023: Fuse the time-domain feature parameters and the frequency-domain feature parameters to generate a sensor feature vector.

[0056] In this embodiment, time-domain feature parameters and frequency-domain feature parameters are fused to form a 64-dimensional sensor feature vector, which is consistent with the dimension of the video feature vector.

[0057] Step S203: Perform structured parsing, knowledge graph construction, and feature vectorization encoding on the historical operation and maintenance data to obtain the knowledge graph feature vector; It should be noted that the historical operation and maintenance data provides prior indications for rule generation and multi-dimensional data association model verification, and achieves structured association through knowledge graph.

[0058] Optionally, step S203 may also include steps S2031 to S2033: Step S2031: Perform structured parsing processing on the historical operation and maintenance data, identify and extract entities corresponding to faults, equipment components, and operation and maintenance measures, as well as the relationships between entities, and generate triplet data based on the relationships between entities; In this embodiment, natural language processing technology is first used to clean and structure text data such as unstructured maintenance work orders, fault logs, and maintenance records in the historical maintenance data. A pre-trained BERT model is then used to perform semantic entity recognition and relation extraction, accurately extracting core entity information such as fault type, equipment component name, and maintenance method. Simultaneously, semantic relationships between entities are mined, including the correspondence between faults and components, faults and maintenance measures, and operating parameters and faults, such as the logical associations like "bearing wear - replace bearing" and "overheating - clean the cooling system." Finally, the scattered unstructured and semi-structured maintenance data is uniformly converted into a standardized (subject-verb-object) knowledge triple format. Exemplary triples include (bearing, fault type, wear) and (vibration peak value, associated fault, bearing wear), achieving a structured and standardized expression of the historical maintenance data.

[0059] Step S2032: Define core entity types and attribute constraints based on the triplet data, and build a knowledge graph by combining expert experience rules; In this embodiment, a knowledge graph for equipment operation and maintenance is built based on the Neo4j graph database. Five core entity types are defined according to structured triplet data, specifically including equipment components, fault types, characteristic parameters, operating conditions, and maintenance measures. Standardized attribute constraints are set for each entity. The fault type entity includes subcategories such as wear, cracks, loosening, and overheating. The characteristic parameter entity is bound to corresponding normal and abnormal threshold ranges. The operating condition entity is associated with operating condition information such as equipment runtime and load level. Furthermore, rules are supplemented and weights are assigned based on the experience of power plant operation and maintenance experts, establishing a strong quantitative correlation between characteristic parameters and fault types. For example, the association weight between "vibration peak > 0.08 mm" and "bearing wear" is set to 0.8, and the association weight between "accumulated equipment runtime > 8000 hours" and equipment aging faults is set to 0.3. This improves the prior operation and maintenance knowledge system of the knowledge graph, achieving a structured modeling of the correlation between equipment faults, operating parameters, and maintenance measures.

[0060] Furthermore, to compensate for the limitations of human expert rules in terms of coverage and difficulty in uncovering latent fault associations, the Apriori (prior association rule) algorithm is used to automatically mine rules from massive equipment operation sample data. The algorithm is configured with a minimum support of 0.05 and a minimum confidence of 0.8, retaining only effective association rules with a sample occurrence frequency of no less than 5% and an association confidence level of no less than 80%. Based on the existing knowledge graph of human prior rules, combined with the parameter-coupled fault associations automatically mined by the algorithm, a complete initial association rule set is obtained, which can accurately mine multi-parameter coupled latent fault rules. An example mining rule is "Vibration peak value > 0.1mm + kurtosis > 4 + infrared temperature > 85℃ → gear crack fault". This effectively supplements the singularity and limitations of human-defined rules, enriches the association system between multi-source features and fault types, and provides more comprehensive prior knowledge support for subsequent model feature association analysis.

[0061] Step S2033: The knowledge graph is vectorized and encoded using a graph embedding algorithm to obtain the knowledge graph feature vector.

[0062] In this embodiment, the TransE graph embedding algorithm is used to perform global vectorization encoding on the constructed knowledge graph. Low-dimensional dense vector mapping is performed on each entity node and relationship in the knowledge graph. The implicit association knowledge of equipment components, fault types, operating parameters, and operation and maintenance rules in the graph is transformed into computable numerical vectors. Finally, a 64-dimensional knowledge graph feature vector is generated, realizing the digital and vectorized expression of operation and maintenance prior knowledge, and providing feature support for operation and maintenance knowledge dimension for subsequent multi-source feature fusion and model association reasoning.

[0063] Step S204: Concatenate the video feature vector, sensor feature vector and knowledge graph feature vector to construct a multi-dimensional training feature vector. Calculate the correlation and correlation weight between each feature node in the multi-dimensional training feature vector using a multi-head graph attention mechanism to complete the weighted representation of the feature nodes and obtain the weighted representation features. In this embodiment, a batch of preprocessed and feature-extracted video feature vectors, sensor feature vectors, and knowledge graph feature vectors are selected and multi-dimensionally concatenated to construct a multi-dimensional training feature vector, which serves as the training input sample for the initial multi-dimensional data association model. Based on the model's built-in dual-attention header graph attention mechanism, the deep association relationship and dynamic association weight of each feature node in the multi-dimensional training feature vector are adaptively and iteratively calculated based on the training samples. Combining the implicit fault association logic mined by the knowledge graph prior rules and prior association rules (Apriori), the coupling association features of equipment operating parameters, component status, and fault types are dynamically strengthened. Differentiated weighted representations are applied to each feature node, fault-sensitive features are selected and highlighted, and invalid noise features are suppressed. Finally, the weighted representation features corresponding to the batch of training samples are output.

[0064] Step S205: Perform multi-head fusion on the weighted representation features, introduce residual connections to optimize gradient propagation, and suppress overfitting of the initial multi-dimensional data association model through a random deactivation mechanism. Iteratively optimize the initial multi-dimensional data association model to obtain the trained multi-dimensional data association model.

[0065] In this embodiment, multiple sets of weighted representation features output by the multi-head attention branch are weighted and summed for multi-head fusion to aggregate effective feature information from multiple dimensions. Residual connections are introduced during training to avoid gradient vanishing, solving the gradient vanishing and gradient explosion problems that easily occur during the training of multi-layer neural networks, and ensuring effective updates of deep network parameters. Simultaneously, a random deactivation mechanism is enabled, with a dropout rate of 0.3, randomly disabling some neuron parameter connections during training, breaking the fixed association patterns of neurons, and effectively suppressing overfitting during model training.

[0066] In this embodiment, the entire dataset of operational samples from all devices is randomly divided into training, validation, and test sets in a 7:2:1 ratio, used for iterative training of model parameters, dynamic tuning of hyperparameters, and final performance verification of the model, respectively. Model training employs the cross-entropy loss function to quantify prediction bias. The loss function is calculated as Loss = -Σy_i log(y'_i), where y_i is the true label and y'_i is the predicted probability. This loss function measures the discrete difference (distribution distance) between the true label distribution and the model's predicted probability distribution, quantifying the degree to which the model's prediction deviates from the standard label. Based on this, the gradient is calculated, providing a basis for iterative optimization through backpropagation. The Adam adaptive optimizer is used for training, configured with an initial learning rate of 0.001 and a first-order moment decay coefficient of 0.9, achieving smooth and efficient updating of network parameters. In this embodiment, the total number of training iterations is set to 100 rounds. After each complete iteration, the model's recognition accuracy and prediction error are verified using a validation set. Based on the validation results, key hyperparameters such as the number of attention heads and dropout rate are dynamically fine-tuned to continuously optimize the multi-source feature association fitting ability of the multi-dimensional data association model. The initial multi-dimensional data association model is iteratively optimized repeatedly until the model's fault recognition accuracy is stable at ≥95% and the prediction error converges to a preset threshold. The iteration is then terminated and all network parameters are solidified, ultimately resulting in a multi-dimensional data association model with strong generalization ability, convergent accuracy, and stable performance.

[0067] By using the methods described above, a multi-dimensional data association model of "video features + sensor data + operation and maintenance history" is constructed. This model integrates algorithms such as deep learning, graph neural networks, and statistical analysis. Through refined data processing and model building throughout the entire process, it enables accurate fault identification, root cause localization, and remaining life prediction, effectively overcoming the limitations of traditional single data source analysis and improving the reliability of decision-making.

[0068] Based on the second embodiment of this application, in the third embodiment of this application, the content that is the same as or similar to that in the second embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 After step S2033, the digital monitoring method further includes steps A1~A2: Step A1: Obtain new operation and maintenance data; Step A2: The newly added operation and maintenance data is supplemented into the constructed knowledge graph through an incremental update algorithm, and the entity associations and association weights in the knowledge graph are updated.

[0069] In this embodiment, new operation and maintenance data throughout the entire lifecycle of the equipment is continuously collected, and data cleaning, sample screening, and accurate labeling of fault categories are completed to form a new sample dataset. For the already constructed equipment operation and maintenance knowledge graph, an incremental iterative update mechanism is adopted to supplement the knowledge graph system with fault entities, component entities, feature parameter entities, and operation and maintenance measure entities corresponding to new fault cases. Simultaneously, the parameter-fault coupling relationship corresponding to the new samples is mined, and the entity association links and association weight parameters within the graph are updated. This fills the gaps in the original knowledge graph's coverage of new fault scenarios and niche fault conditions, achieving dynamic iterative expansion of the operation and maintenance knowledge system.

[0070] Simultaneously, an incremental learning algorithm is employed, which combines parameter freezing with local fine-tuning. This fixes the parameters of the input and feature fusion layers of the multi-dimensional data association model, while only fine-tuning the network parameters of the graph attention and output layers. This significantly reduces the computational overhead of iterative training of the multi-dimensional data association model. In this embodiment, the incremental update trigger condition is set to perform an incremental iterative update of the knowledge graph and model parameters every 50 newly added labeled fault samples. This ensures that the multi-dimensional data association model continuously adapts to new fault scenarios, with parameter update amplitude controlled within ±10%, guaranteeing the stability of the multi-dimensional data association model.

[0071] Through the methods described above, newly added operation and maintenance data, after being manually reviewed, is automatically added to the knowledge graph through an incremental update algorithm, updating the entity relationship weights and ensuring the timeliness and completeness of the knowledge graph.

[0072] Based on the first embodiment of this application, in the fourth embodiment of this application, the content that is the same as or similar to that in the first embodiment described above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 After step S20, the digital monitoring method further includes steps B1 to B4: Step B1: Based on the probability distribution of the fault type and the prediction results of the remaining life of the equipment, obtain the abnormal monitoring results corresponding to the video data, sensor data and operation and maintenance history data; In this embodiment, based on the probability distribution of fault types and the prediction results of remaining equipment lifespan, the multi-source heterogeneous data of the original input are reversed and matched to extract the equipment anomaly monitoring results corresponding to the video visual monitoring dimension, sensor operating parameter dimension, and operation and maintenance history prior knowledge dimension.

[0073] Step B2: Weighted fusion scoring is performed on the credibility weights corresponding to each anomaly monitoring result to obtain the weighted score corresponding to each anomaly monitoring result; In this embodiment, a multi-source data credibility assessment system is constructed, and credibility weights are assigned to three types of monitoring data. The credibility weight for sensor data is 0.4, for video feature data it is 0.3, and for historical maintenance knowledge data it is 0.3. For the anomaly monitoring results output from each dimension, a base score of 0-1 is assigned to each dimension based on the degree of anomaly matching. This score is then combined with the corresponding credibility weight for weighted fusion calculation to obtain a comprehensive anomaly weighted score for the multi-source data.

[0074] The specific judgment logic is as follows: when a suspected anomaly is identified by a single data source, it is initially judged as a valid anomaly only if the weighted total score after the fusion of multi-source data is ≥0.5; otherwise, it is directly judged as a suspected false alarm sample.

[0075] For example, if only video data identifies a suspected crack on the device surface, the basic score of the video feature is 0.3. The sensor vibration and temperature parameters are normal and have a basic score of 0. There are no similar fault records in the maintenance history and the basic score is 0. The comprehensive weighted total score is 0.3, which is less than the 0.5 threshold. This is determined to be a false alarm caused by image shadows or surface stains. If the sensor detects an abnormal vibration peak and has a basic score of 1, corresponding to a score of 0.4, the video detects a component temperature that is too high and has a basic score of 0.5, corresponding to a score of 0.15. There are similar vibration fault records in the maintenance history and have a basic score of 1, corresponding to a score of 0.3. The comprehensive weighted total score is 0.85, which is greater than the 0.5 threshold. This is determined to be a real and valid equipment anomaly, triggering the fault warning mechanism.

[0076] Step B3: Compare the weighted score with the dynamic anomaly threshold, and determine the fault anomaly and false alarm samples based on the comparison result. The dynamic anomaly threshold is obtained by statistical analysis of the equipment's historical normal operation data and adaptive adjustment of the equipment's real-time operating conditions. In this embodiment, the shortcomings of traditional fixed thresholds, which are susceptible to false alarms due to fluctuations in equipment load, changes in ambient temperature, and switching of operating conditions, are eliminated. Instead, a two-dimensional dynamic threshold adaptive adjustment mechanism is constructed. For sensor numerical characteristic parameters, statistical analysis is performed on data from the equipment's stable normal operation over the past 72 hours. The "mean + 2.5 times the standard deviation" is used as the dynamic judgment threshold, and the threshold parameter is updated every 24 hours to adapt to the dynamic changes in equipment operating conditions in real time.

[0077] For example, when the average value of the peak vibration of the equipment bearing is 0.03 mm and the standard deviation is 0.02 mm during normal operation, the dynamic threshold is calculated to be 0.08 mm; when the equipment operates under high load for a long time, causing the operating conditions to deviate and the average value is updated to 0.04 mm, the dynamic threshold is automatically adapted and adjusted to 0.09 mm to ensure that the threshold is accurately matched with the actual operating conditions.

[0078] Based on video features, a normal equipment image library and a fault feature image library are constructed. A cosine similarity algorithm is used to calculate the feature similarity between suspected fault areas and standard fault images. A similarity benchmark threshold of 0.7 is set. When the image feature similarity is less than 0.7, it is judged as a suspected false alarm caused by image interference and included in the false alarm investigation. Through dual verification of dynamic threshold and weighted score, the system accurately distinguishes between actual equipment faults and false alarms caused by environmental or equipment disturbances.

[0079] Step B4: Build a false alarm case library based on the fault anomalies and false alarm samples, label the false alarm reasons for the false alarm samples, and perform statistical analysis on the labeled false alarm samples to optimize the attention weight and the dynamic anomaly threshold in the multi-dimensional data association model.

[0080] In this embodiment, real fault samples and false alarm samples obtained through cross-validation and dynamic threshold verification are categorized and collected to build a standardized equipment monitoring false alarm case library. The causes of false alarms are labeled for all false alarm samples, such as image interference, instantaneous sensor fluctuations, and sudden changes in operating conditions. Statistical analysis of false alarm cases in the case library is performed quarterly to optimize the model's attention weights and threshold parameters. For example, if false alarms caused by image shadows account for 60%, the weights of features such as "edge sharpness" and "grayscale uniformity" in the video features are strengthened to reduce interference from shadow areas and further compress the false alarm space.

[0081] Through the above embodiments, a triple false alarm filtering mechanism of "multi-feature cross-validation + dynamic operating condition threshold adaptive adjustment + false alarm case feedback iteration" is adopted to completely solve the problems of high false alarms and high false negatives caused by traditional fixed threshold and single data source monitoring, and strictly control the false alarm rate of equipment fault monitoring to below 2%. At the same time, relying on the case library, the model attention weight and judgment threshold are continuously iterated and optimized to ensure that the model always adapts to the changes in operating conditions and new fault scenarios throughout the entire life cycle of the equipment, which greatly improves the accuracy, stability and robustness of digital power plant equipment fault monitoring, and further strengthens the system's proactive intelligent operation and maintenance capabilities.

[0082] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the digital monitoring method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0083] This application also provides a digital monitoring system; please refer to [reference needed]. Figure 5 The digital monitoring system includes: The acquisition module 10 is used to acquire multi-source data of device operation, including one or more of video data, sensor data, and historical operation and maintenance data; The inference module 20 is used to input the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing, and output the probability distribution of fault types and the prediction results of the remaining life of the equipment. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction and association training on the multi-source data.

[0084] The digital monitoring system provided in this application, employing the digital monitoring method described in the above embodiments, can solve the technical problem that existing technologies struggle to meet the high-precision, high-real-time, and intelligent operation and maintenance requirements of digital power plants. Compared with existing technologies, the beneficial effects of the digital monitoring system provided in this application are the same as those of the digital monitoring method described in the above embodiments, and other technical features of the digital monitoring system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0085] This application provides a digital monitoring device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the digital monitoring method in Embodiment 1 above.

[0086] The following is for reference. Figure 6 The diagram illustrates a structural schematic of a digital monitoring device suitable for implementing embodiments of this application. The digital monitoring device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The digital monitoring device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0087] like Figure 6As shown, the digital monitoring device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the digital monitoring device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the digital monitoring device to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows digital monitoring devices with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0088] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0089] The digital monitoring equipment provided in this application, employing the digital monitoring method described in the above embodiments, can solve the technical problem that existing technologies struggle to meet the high-precision, high-real-time, and intelligent operation and maintenance requirements of digital power plants. Compared with existing technologies, the beneficial effects of the digital monitoring equipment provided in this application are the same as those of the digital monitoring method described in the above embodiments, and other technical features of this digital monitoring equipment are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0090] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0092] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the digital monitoring method described in the above embodiments.

[0093] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0094] The aforementioned computer-readable storage medium may be included in the digital monitoring device; or it may exist independently and not be assembled into the digital monitoring device.

[0095] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the digital monitoring device, the digital monitoring device causes the following actions: to collect multi-source data of the device operation, wherein the multi-source data includes one or more of video data, sensor data, and historical operation and maintenance data; to input the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing; and to output a fault type probability distribution and a prediction result of the device's remaining lifespan. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction, and association training on the multi-source data.

[0096] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0097] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0098] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0099] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the above-described digital monitoring method. This solves the technical problem that existing technologies struggle to meet the high-precision, high-real-time, and intelligent operation and maintenance requirements of digital power plants. Compared with existing technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the digital monitoring method provided in the above embodiments, and will not be elaborated upon here.

[0100] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the digital monitoring method described above.

[0101] The computer program product provided in this application can solve the technical problem that existing technologies cannot meet the high-precision, high-real-time, and intelligent operation and maintenance requirements of digital power plants. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the digital monitoring method provided in the above embodiments, and will not be repeated here.

[0102] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A digital monitoring method, characterized in that, The digital monitoring method includes: Collect multi-source data on the operation of the acquisition equipment, wherein the multi-source data includes one or more of video data, sensor data, and historical operation and maintenance data; The multi-source data is input into a pre-trained multi-dimensional data association model for feature association parsing, and the outputs the probability distribution of fault types and the prediction results of the remaining life of the equipment. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction and association training on the multi-source data.

2. The digital monitoring method as described in claim 1, characterized in that, The step of inputting the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing, and outputting the fault type probability distribution and equipment remaining life prediction results includes: Based on the video data, sensor data, and historical operation and maintenance data, the corresponding video feature vectors, sensor feature vectors, and knowledge graph feature vectors are extracted respectively. The video feature vector, sensor feature vector, and knowledge graph feature vector are concatenated into a multi-dimensional input feature vector, which is then input into a pre-trained multi-dimensional data association model. In the pre-trained multi-dimensional data association model, the association weight coefficients between each feature node in the multi-dimensional input feature vector are calculated through a multi-head graph attention mechanism; Based on the aforementioned correlation weight coefficients, each feature node is weighted and represented to enhance the expression of the correlation features between equipment operating parameters and equipment components, thereby obtaining weighted representation features. The weighted representation features are weighted, summed, and fused to generate a fused feature vector; Based on the fused feature vector, the probability distribution of fault types corresponding to various faults is calculated using an activation function, and the equipment fault status is determined by combining the preset probability threshold. The fused feature vector is input into a pre-trained Long Short-Term Memory (LSTM) network model for lifetime prediction mapping to obtain the device's remaining lifetime prediction result.

3. The digital monitoring method as described in claim 1, characterized in that, Before the step of inputting the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing and outputting the fault type probability distribution and equipment remaining life prediction results, the following steps are included: Frame sampling preprocessing and feature layer extraction are performed on the video data to obtain video feature vectors; Sensor data undergoes data preprocessing and time-frequency domain feature extraction to obtain sensor feature vectors; The historical operation and maintenance data is subjected to structured parsing, knowledge graph construction, and feature vectorization encoding to obtain knowledge graph feature vectors; The video feature vector, sensor feature vector, and knowledge graph feature vector are concatenated to construct a multi-dimensional training feature vector. The association relationship and association weight between each feature node in the multi-dimensional training feature vector are calculated through a multi-head graph attention mechanism to complete the weighted representation of the feature nodes and obtain the weighted representation features. The weighted representation features are fused from multiple heads, residual connections are introduced to optimize gradient propagation, and a random deactivation mechanism is used to suppress overfitting of the initial multidimensional data association model. The initial multidimensional data association model is iteratively optimized to obtain the trained multidimensional data association model.

4. The digital monitoring method as described in claim 3, characterized in that, The steps of performing frame sampling preprocessing and feature layer extraction on the video data to obtain the video feature vector include: The video data is sampled frame by frame, and the sampled video data is preprocessed by denoising, contrast optimization, infrared temperature calibration and pixel normalization to obtain preprocessed video data. Based on a pre-trained residual convolutional neural network model, hierarchical feature extraction is performed on the preprocessed video data to obtain multi-level feature vectors. The multi-level feature vectors are compressed in dimension using global average pooling to obtain the initial video feature vectors. A channel attention mechanism is introduced to assign weights to the initial video feature vector, resulting in the final video feature vector.

5. The digital monitoring method as described in claim 3, characterized in that, The steps of performing data preprocessing and time-frequency domain feature extraction on sensor data to obtain sensor feature vectors include: The sensor data is subjected to outlier removal, missing value supplementation and filtering noise reduction to obtain preprocessed sensor data; Based on the preprocessed sensor data, time-domain feature parameters are extracted from the time-domain dimension, and frequency-domain feature parameters are extracted from the frequency-domain dimension through signal transformation. The time-domain feature parameters and the frequency-domain feature parameters are fused to generate a sensor feature vector.

6. The digital monitoring method as described in claim 3, characterized in that, The steps of performing structured parsing, knowledge graph construction, and feature vectorization encoding on historical operation and maintenance data to obtain knowledge graph feature vectors include: The historical operation and maintenance data is subjected to structured parsing processing to identify and extract entities corresponding to faults, equipment components, and operation and maintenance measures, as well as the relationships between entities. Based on the relationships between entities, triplet data is generated. Based on the triplet data, the core entity types and attribute constraints are defined, and a knowledge graph is built by combining expert experience rules. The knowledge graph is vectorized by using a graph embedding algorithm to obtain the knowledge graph feature vector.

7. The digital monitoring method as described in claim 6, characterized in that, After the step of vectorizing the knowledge graph using a graph embedding algorithm to obtain the knowledge graph feature vector, the following steps are included: Get new operation and maintenance data; The newly added operation and maintenance data is supplemented into the constructed knowledge graph through an incremental update algorithm, thereby updating the entity relationships and association weights in the knowledge graph.

8. The digital monitoring method as described in claim 1, characterized in that, After the step of inputting the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing and outputting the fault type probability distribution and equipment remaining life prediction results, the following steps are included: Based on the probability distribution of the fault types and the prediction results of the remaining lifespan of the equipment, the abnormal monitoring results corresponding to the video data, sensor data and operation and maintenance history data are obtained. The credibility weights of each anomaly monitoring result are weighted and fused to obtain the weighted score for each anomaly monitoring result. The weighted score is compared with the dynamic anomaly threshold, and the fault anomaly and false alarm samples are determined based on the comparison results. The dynamic anomaly threshold is obtained by statistical analysis of the equipment's historical normal operation data and adaptive adjustment of the equipment's real-time operating conditions. A false alarm case library is built based on the aforementioned fault anomalies and false alarm samples. The false alarm samples are labeled with false alarm causes, and statistical analysis is performed on the labeled false alarm samples to optimize the attention weight and the dynamic anomaly threshold in the multi-dimensional data association model.

9. A digital monitoring system, characterized in that, The digital monitoring includes: The acquisition module is used to collect multi-source data of equipment operation, including one or more of video data, sensor data, and historical operation and maintenance data. The inference module is used to input the multi-source data into a pre-trained multi-dimensional data association model for feature association parsing, and output the probability distribution of fault types and the prediction results of the remaining life of the equipment. The multi-dimensional data association model is obtained by performing data preprocessing, feature extraction and association training on the multi-source data.

10. A digital monitoring device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the digital monitoring method as described in any one of claims 1 to 7.

Images