Vector and time series database driven industrial cognitive analysis method and system

Abstract

This invention relates to the field of data processing technology, and in particular to an industrial cognitive analysis method and system driven by vector and time-series databases. The method constructs a two-layer architecture of knowledge graph and variational manifold space. Through manifold embedding, time-series data is mapped into latent state vectors containing deterministic and fluctuating characteristics. Combined with semantic vectors, isomorphic projection based on drift gradient is performed, and then a dynamic Bayesian graph neural network is used to deduce the fault path. This method solves the problem of semantic gap and poor noise resistance between continuous monitoring data and discrete knowledge entities. While accommodating the uncertainty of random disturbances in the industrial environment, it achieves accurate semantic alignment of multimodal data, can keenly capture the gradual trend of state change, and dynamically restores the fault evolution path that conforms to logical causality, significantly improving the robustness and interpretability of diagnosis.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an industrial cognitive analysis method and system driven by vector and time-series databases. Background Technology

[0002] With the rapid development of industrial internet and intelligent manufacturing technologies, the operation and maintenance of modern industrial equipment increasingly rely on the in-depth mining and analysis of massive multimodal data. On the one hand, in actual industrial production sites, various sensors continuously generate high-frequency, continuous time-series monitoring data, which contain the real-time physical state and fluctuation patterns of equipment operation. On the other hand, in the long-term process of equipment design and operation and maintenance, a large amount of unstructured semantic knowledge has been accumulated, such as equipment maintenance manuals, expert experience bases, and fault case descriptions. This knowledge usually exists in the form of discrete entities and their causal relationships, which are the key basis for fault diagnosis and decision-making.

[0003] However, since time-series data exhibits continuous numerical fluctuations while industrial knowledge is represented by discrete symbolic entities, the two exist in completely different feature modes, resulting in a natural semantic gap. More specifically, in complex industrial environments, sensor data is often accompanied by random disturbances, noise, and uncertainty in state. Therefore, the industry urgently needs a new analytical method that can not only achieve semantic alignment between multimodal data and knowledge nodes while accommodating data uncertainty, but also dynamically deduce the occurrence mechanism and evolution path of faults in the knowledge graph based on this alignment relationship, thereby providing interpretable analytical results for the precise operation and maintenance of equipment. Summary of the Invention

[0004] The main objective of this invention is to provide an industrial cognitive analysis method driven by vector and time-series databases, which aims to solve the problem that existing industrial retrieval and analysis methods are unable to achieve efficient integration and dynamic interaction between continuous time-series data and discrete industrial knowledge.

[0005] To achieve the above objectives, this invention provides an industrial cognitive analysis method driven by vector and time-series databases, the method comprising the following steps:

[0006] A two-layer spatial architecture for industrial knowledge is constructed, comprising an industrial knowledge graph and a temporal variational manifold space. The industrial knowledge graph is used to store discrete entities and their causal relationships, while the temporal variational manifold space is used to characterize the distribution patterns of continuous temporal data.

[0007] Acquire multimodal data from the industrial site, including time-series data of industrial equipment operation and unstructured semantic data describing the state of industrial equipment, including text data and / or image data;

[0008] Manifold embedding processing is performed on the time series data to map the time series data onto the time series variational manifold space to obtain a latent state vector containing state deterministic features and uncertain fluctuation features.

[0009] Semantic feature encoding is performed on the unstructured semantic data to transform the text data and / or image data into semantic vectors representing current working condition information;

[0010] The activation energies of the latent state vector and the semantic vector at each graph node corresponding to the discrete entity are obtained by isomorphic projection operation based on drift gradient and semantic alignment.

[0011] Using the activation energy as the initial state, a dynamic Bayesian graph neural network is run along the causal relationship of the discrete entities on the industrial knowledge graph to perform probability propagation in order to obtain the target fault discrete entity of the current equipment state and the evolution path of the target fault discrete entity.

[0012] The target fault discrete entity and its evolution path are output as analysis results.

[0013] To achieve the above objectives, the present invention also provides a database-driven industrial knowledge analysis system, the system comprising:

[0014] A two-layer space construction module is used to construct a two-layer space architecture for industrial knowledge. The two-layer space architecture includes an industrial knowledge graph and a temporal variational manifold space. The industrial knowledge graph is used to store discrete entities and their causal relationships, and the temporal variational manifold space is used to characterize the distribution patterns of continuous temporal data.

[0015] A multimodal data acquisition module is used to acquire multimodal data from the industrial site. The multimodal data includes time-series data of industrial equipment operation and unstructured semantic data describing the state of industrial equipment. The unstructured semantic data includes text data and / or image data.

[0016] The manifold embedding processing module is used to perform manifold embedding processing on the time series data, mapping the time series data to the time series variational manifold space to obtain a latent state vector containing state deterministic features and uncertain fluctuation features.

[0017] The semantic feature encoding module is used to encode the unstructured semantic data into semantic features, transforming the text data and / or image data into semantic vectors that represent current working condition information.

[0018] The isomorphic projection calculation module is used to obtain the activation energy of the latent state vector and the semantic vector at each graph node corresponding to the discrete entity based on the isomorphic projection operation based on drift gradient and semantic alignment.

[0019] The dynamic reasoning module is used to run a dynamic Bayesian graph neural network to perform probability propagation along the causal relationship of the discrete entities on the industrial knowledge graph, with the activation energy as the initial state, in order to obtain the target fault discrete entity of the current equipment state and the evolution path of the target fault discrete entity.

[0020] The result output module is used to output the target fault discrete entity and the evolution path of the target fault discrete entity as analysis results.

[0021] To achieve the above objectives, the present invention also provides a computer device including a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program.

[0022] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing a computer program, wherein a processor executes the computer program.

[0023] The embodiments of the present invention propose,

[0024] This invention constructs a two-layer underlying architecture comprising an industrial knowledge graph and a temporal variational manifold space. It utilizes manifold embedding and variational reasoning techniques to map continuous and noisy industrial time-series monitoring data into potential state vectors exhibiting both deterministic trends and uncertain fluctuations. Furthermore, it combines semantic encoding of unstructured text or image data with isomorphic projection operations based on drift gradients and semantic alignment to calculate the activation energy of discrete entities. This energy then drives a dynamic Bayesian graph neural network to perform probability propagation and path optimization on the causal topology of the knowledge graph. This effectively solves the technical problems in existing technologies where a natural semantic gap exists between continuous time-series monitoring data and discrete industrial knowledge entities, and where accurate dynamic association and logical reasoning are difficult to achieve in noisy and uncertain industrial environments.

[0025] It achieves deep fusion and precise semantic alignment of multimodal heterogeneous data while effectively accommodating random data disturbances. It can not only automatically filter non-faulty random noise through variational manifold space, giving the system strong robustness, but also use drift gradient projection to keenly capture the early gradual change trend of equipment status and restore the logical causal fault evolution path through dynamic Bayesian inference. This significantly improves the accuracy and interpretability of fault diagnosis of complex industrial equipment, and realizes the intelligent leap from passive post-event diagnosis to proactive predictive maintenance. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the method in Embodiment 1 of the present invention;

[0027] Figure 2 This is a structural block diagram of the system in Embodiment 7 of the present invention;

[0028] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0031] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0032] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0033] Example 1

[0034] As attached Figure 1 As shown, this embodiment provides an industrial cognitive analysis method driven by vector and time-series databases. The method includes the following steps:

[0035] A two-layer spatial architecture for industrial knowledge is constructed, comprising an industrial knowledge graph and a temporal variational manifold space. The industrial knowledge graph is used to store discrete entities and their causal relationships, while the temporal variational manifold space is used to characterize the distribution patterns of continuous temporal data.

[0036] Acquire multimodal data from the industrial site, including time-series data of industrial equipment operation and unstructured semantic data describing the state of industrial equipment, including text data and / or image data;

[0037] Manifold embedding processing is performed on the time series data to map the time series data onto the time series variational manifold space to obtain a latent state vector containing state deterministic features and uncertain fluctuation features.

[0038] Semantic feature encoding is performed on the unstructured semantic data to transform the text data and / or image data into semantic vectors representing current working condition information;

[0039] The activation energies of the latent state vector and the semantic vector at each graph node corresponding to the discrete entity are obtained by isomorphic projection operation based on drift gradient and semantic alignment.

[0040] Using the activation energy as the initial state, a dynamic Bayesian graph neural network is run along the causal relationship of the discrete entities on the industrial knowledge graph to perform probability propagation in order to obtain the target fault discrete entity of the current equipment state and the evolution path of the target fault discrete entity.

[0041] The target fault discrete entity and its evolution path are output as analysis results.

[0042] In the fields of industrial internet and intelligent manufacturing, the deep mining of multimodal data faces a core challenge: the semantic gap between continuous time-series monitoring data and discrete industrial knowledge. Specifically, high-frequency time-series data collected by sensors exhibits continuous numerical fluctuations, while unstructured semantic data such as equipment maintenance manuals and expert experience bases exist in the form of discrete symbolic entities, resulting in significant differences in their characteristic modes. At the same time, random disturbances and noise interference in industrial sites exacerbate the uncertainty of data states, making it difficult for existing technologies to achieve semantic alignment of multimodal data to knowledge nodes while accommodating this uncertainty. This hinders the process of dynamically deducing fault occurrence mechanisms and evolution paths in knowledge graphs based on alignment relationships, ultimately affecting the reliability and interpretability of fault diagnosis results.

[0043] To address the aforementioned issues, this embodiment provides an industrial cognitive analysis method driven by vector and time-series databases. By constructing a two-layer spatial architecture for industrial knowledge, it achieves deep fusion and dynamic reasoning between continuous data and discrete knowledge. It is understood that the method in this embodiment runs on specific computer equipment or server clusters. First, it requires constructing and initializing the underlying two-layer spatial architecture. This architecture includes an industrial knowledge graph for storing discrete entities and their causal relationships, and a time-series variational manifold space for representing the distribution patterns of continuous time-series data. The industrial knowledge graph is typically deployed in a graph database, where discrete entities correspond to physical components, failure modes, or operational events in an industrial scenario, while causal relationships define the topological structure of fault propagation or logical associations between these entities. The time-series variational manifold space is deployed in a high-dimensional vector database; its essence is a trained manifold topology structure used to accommodate high-dimensional distributions mapped from continuous time-series data.

[0044] It is also understandable that once the two-layer spatial architecture is built, the processor begins to acquire multimodal data from the industrial site in real time. This data includes both high-frequency time-series monitoring data collected by sensor networks, such as motor vibration acceleration and bearing temperature curves, and unstructured semantic data describing the current operating conditions of the equipment, such as text logs filled out by maintenance personnel or images taken by infrared thermal imagers. For the acquired time-series data, the system does not directly compare it with the raw values, because noise in the industrial environment can easily cause spurious fluctuations in the values. Instead, the processor performs manifold embedding processing, using a pre-built multimodal encoder to extract features from the time-series data within a sliding time window, and introduces a variational inference mechanism to map the data into a probability distribution region in the time-series variational manifold space, thereby generating a latent state vector. This process is calculated using the following latent state vector generation model, i.e., Expression 1: In the formula, The vector representing the latent state generated at time t contains both deterministic and uncertain fluctuation characteristics and is a high-dimensional vector located in the temporal variational manifold space.

[0045] The matrix representing the raw time series data within the current time window at time t is typically composed of a continuous signal sequence collected by sensors in the industrial field.

[0046] This represents the set of all learnable neural network weights and bias parameters in the preset multimodal encoder;

[0047] This indicates that the encoder network consists of parameters. The determined mean mapping function is used to output the mean vector representing the current center position of the data distribution;

[0048] This indicates that the encoder network consists of parameters. The determined log-variance mapping function is used to output the log-variance vector that characterizes the degree of dispersion of the current data distribution;

[0049] This represents auxiliary noise variables, which are random noise variables sampled from a standard normal distribution;

[0050] The Hadamard product represents the operation of multiplying corresponding elements of two vectors one by one.

[0051] This represents an exponential function operation with the natural constant e as the base, used to reduce the logarithmic variance to the standard deviation. In special scenarios, such as when a sensor experiences instantaneous changes due to electromagnetic interference, traditional point mapping methods may misinterpret this as a sudden change in device state. The above expression, by introducing a variance term and random noise sampling, forces the model to learn the distribution area of ​​the data in the manifold space rather than a single coordinate point. When the data noise is high or in an unknown transitional state, the value of the variance function output will increase, thus widening the spatial range covered by the potential state vector. This processing gives the system strong robustness, enabling it to automatically filter high-frequency random noise and capture only statistically significant state drifts, thereby achieving accurate feature representation in uncertain environments.

[0052] Furthermore, for the acquired unstructured semantic data, the processor performs semantic feature encoding. If the input is text data, a natural language processing model is used to extract key entities and related information; if the input is image data, a convolutional neural network is used to extract visual features, and finally it is transformed into a semantic vector representing the current working condition information. This vector is in the same high-dimensional vector space dimension as the potential state vector.

[0053] After feature mapping is completed, the core step is to associate vectors in the continuous space with nodes in the discrete graph. The processor performs isomorphic projection operations based on drift gradients and semantic alignment to calculate the activation energy of the latent state vector and semantic vector on the graph nodes corresponding to each discrete entity. This process no longer relies on a single similarity matching, but integrates physical distance, semantic consistency, and state evolution trends. The calculation of activation energy is based on the following isomorphic projection equation, i.e., Expression 2: In the formula, This represents the activation energy for the k-th discrete entity in the industrial knowledge graph;

[0054] This represents the static baseline eigenvector corresponding to the k-th discrete entity in the manifold space;

[0055] This represents the weighted Mahalanobis distance, which uses the inverse of the covariance matrix of historical data to eliminate variable correlation.

[0056] A semantic vector representing the current operating condition;

[0057] The time derivative of the potential state vector as a function of time, i.e., the drift gradient vector;

[0058] This represents the principal feature evolution direction vector of the k-th discrete entity;

[0059] ω1, ω2, ω3 represent the adaptive weight coefficients of each matching item;

[0060] Let represent the hyperbolic tangent activation function. In the above expression 2, the first term of the weighted fusion measures the static matching degree of the physical state using Mahalanobis distance, the second term measures the logical consistency of the semantic description, and the third term uses drift gradient projection to capture trends. In the early stages of industrial failures, the state data may not have exceeded the threshold (i.e., the first term distance is still close), but its rate of change and direction (drift gradient) may already point to a specific failure mode. By projecting the gradient onto the principal feature direction of the entity, this gradual failure trend can be keenly identified, so that the activation energy not only reflects what the current state is, but also predicts what the state is becoming, thereby significantly improving the early warning capability for evolutionary failures.

[0061] Finally, using the calculated activation energy as the initial state, the processor runs a dynamic Bayesian graph neural network along the causal relationships of discrete entities on the industrial knowledge graph. This process simulates the thought process of experts reasoning based on phenomena and experience. The system calculates the posterior propagation probability based on the activation energy and graph structure, thereby deriving the evolution path of the fault. The calculation of the posterior propagation probability follows the following Bayesian inference formula, i.e., Expression 3: In the formula, This represents the posterior propagation probability of a fault spreading from parent node x to child node y.

[0062] β represents the adaptive balance factor adjusted based on data confidence.

[0063] This represents the real-time activation energy of child node y;

[0064] This represents the activation energy normalization factor for all nodes within a node's neighborhood.

[0065] This represents the predefined causal transition probability from node x to node y in a knowledge graph;

[0066] This represents the semantic association attention coefficient between node x and node y. In Expression 3 above, the first term represents data evidence, entirely dependent on real-time sensor and semantic feedback; the second term represents prior knowledge, dependent on causal logic in the expert database. The adaptive balancing factor β is dynamically adjusted based on the signal-to-noise ratio of the data: when the sensor data quality is high, β increases, and inference relies on real-time observation; when data is missing or semantically ambiguous, β decreases, and inference automatically reverts to relying on expert experience. This mechanism effectively solves the deficiency of a single data-driven model that cannot even provide a guess when the sensor fails, ensuring that the system can still output the most logical fault evolution path under extreme conditions, and ultimately determining the endpoint of the path with the highest cumulative probability as the target fault discrete entity and outputting it.

[0067] In summary, this embodiment constructs a two-layer underlying architecture comprising an industrial knowledge graph and a temporal variational manifold space. It utilizes manifold embedding and variational inference techniques to map continuous and noisy industrial time-series monitoring data into potential state vectors exhibiting both deterministic trends and uncertain fluctuations. Combined with semantic encoding of unstructured text or image data, it performs isomorphic projection operations based on drift gradients and semantic alignment to calculate the activation energy of discrete entities. This energy then drives the operation of a dynamic Bayesian graph neural network on the causal topology of the knowledge graph for probability propagation and path optimization. This addresses the technical challenges of existing technologies, such as the inherent semantic gap between continuous time-series monitoring data and discrete industrial knowledge entities, and the difficulty in achieving accurate dynamic association and logical reasoning in noisy and uncertain industrial environments. It achieves deep fusion and accurate semantic alignment of multimodal heterogeneous data while effectively accommodating random data disturbances. Furthermore, it can keenly capture the gradual drift trends of equipment states and dynamically deduce logically causal fault evolution paths, thereby significantly improving the robustness, accuracy, and interpretability of fault diagnosis for complex industrial equipment.

[0068] In this embodiment, the step of performing manifold embedding processing on the time-series data, mapping the time-series data to the time-series variational manifold space to obtain a latent state vector containing both deterministic and uncertain fluctuation features, specifically includes:

[0069] Feature extraction is performed on the time series data within the current time window based on a preset multimodal encoder;

[0070] The mean vector and variance vector, which characterize the current data distribution, are obtained from the output of the multimodal encoder according to the variational inference mechanism.

[0071] The latent state vector is obtained by combining reparameterization with Gaussian noise sampling, and then by using the mean vector and variance vector, so that the latent state vector can accommodate random disturbances and noise uncertainties in the industrial environment.

[0072] It should be noted that in the above expression one, the anchor point or main trend of equipment operation is reflected. For example, when early wear occurs in a bearing, even if there is a lot of noise in the signal, the energy centroid (mean vector) of its vibration signal will undergo a definite displacement in the manifold space. This feature ensures that the model can accurately identify the type of fault. Secondly, in the formula... The term represents the characteristics of uncertainty and fluctuation.

[0073] This embodiment features adaptive logic specifically designed for industrial scenarios. In actual operating conditions, sensors may experience non-fault-related reading fluctuations due to factors such as base loosening, temperature drift, or transient electromagnetic pulses. If the model only learns the mean, these fluctuations are easily misinterpreted as fault characteristics. By introducing a logarithmic variance term and restoring it to the standard deviation, the model effectively constructs a logarithmic variance term centered at μ in the manifold space. A probability cloud with radius .

[0074] When input data When exhibiting high regularity and stability, the log-variance of network predictions tends to negative infinity (i.e., the variance approaches 0). The data tends to approach the mean, exhibiting high confidence; however, when the input data contains a lot of noise or is in some unknown transitional mode, the network automatically predicts a larger variance, resulting in a less reliable output. This design allows for greater spatial coverage. An additional technical benefit is that in subsequent projection calculations, if the variance of a state vector is large (i.e., the "cloud" is large), its matching degree with a specific fault node in the knowledge graph will be diluted or its weight reduced. This naturally suppresses false alarms caused by sensor noise. This reparameterized generation method not only makes the entire calculation process mathematically differentiable (supporting backpropagation training), but also physically endows the system with robust analytical capabilities similar to those of a senior engineer, focusing on the major issues and investigating those in doubt.

[0075] Example 2:

[0076] In this embodiment, the semantic feature encoding of the unstructured semantic data specifically includes:

[0077] When the unstructured semantic data contains text data, a preset language processing model is used to extract text keywords and their relationships, and these are mapped into a high-dimensional vector as the semantic vector.

[0078] When the unstructured semantic data includes image data, visual features are extracted from the image data using a convolutional neural network, and these visual features are aligned into the semantic space. The resulting high-dimensional vector is then defined as the semantic vector.

[0079] It should be noted that in practical applications, unstructured semantic data typically refers to text and image data that cannot be directly stored in relational databases. Examples include handwritten troubleshooting logs by equipment maintenance personnel, operation manuals in digital delivery documents, thermal distribution maps of equipment taken by infrared thermal imagers, or internal flaw detection images collected by industrial endoscopes. These data contain a wealth of explicit descriptions and implicit clues about equipment operating conditions. However, due to their distinct modalities compared to time-series monitoring data, a significant semantic gap exists. Therefore, this embodiment designs a dual-channel semantic feature encoding mechanism to map heterogeneous text and image data into the same high-dimensional semantic space, generating semantic vectors with the same dimension as the latent state vectors.

[0080] For unstructured semantic data containing text, the system invokes a pre-defined language processing model to perform deep semantic parsing. The preferred language processing model is a pre-trained model based on the Transformer architecture (such as the industrial version of BERT or RoBERTa), which possesses powerful contextual understanding capabilities. In practice, the processor first segments and embeds vectorized text data (e.g., "No. 3 pump unit vibrates abnormally, suspected bearing cage breakage") into words. Then, it dynamically captures key entities in the text (e.g., "pump unit," "bearing") and their relationships (e.g., "abnormal vibration," "breakage") through a multi-head self-attention mechanism. To aggregate this discrete word information into a global semantic representation, this embodiment uses the following text semantic vector generation formula for calculation: In the formula, The generated text semantic vector is a highly compressed and abstracted representation of the current input text, which is directly used for subsequent isomorphic projection calculations.

[0081] L represents the length of the input text sequence, which is the total number of tokens after word segmentation;

[0082] This represents the contextual hidden layer vector of the i-th word output by the language processing model. This vector not only contains the meaning of the word itself, but also incorporates its contextual position information in the sentence.

[0083] This represents the attention weight coefficient of the i-th word, which is automatically learned by the model based on the importance of the word's contribution to the overall semantics, and is used to highlight keywords such as faulty entities;

[0084] This represents the semantic projection matrix, which is used to map the aggregated feature space to a space with the same dimension as the temporal latent state vector.

[0085] This represents the bias vector of the projection layer, used to adjust the reference point of the vector space;

[0086] This represents the hyperbolic tangent activation function, used to normalize the numerical range of the output vector to the interval [-1, 1] to prevent numerical explosion.

[0087] It should be noted that industrial logs are often filled with a large amount of useless information (such as dates and irrelevant interjections), while the truly core information (keywords and their associations) only constitutes a small portion. The weighted summation term in the formula is essentially an attention aggregation process; the model assigns higher weights to key technical terms such as "bearing" and "fracture," while suppressing the weights of meaningless words. In this way, the generated... It's no longer a simple bag-of-words overlay, but rather a precise capture of causal logic and entity relationships within the text. (Projection matrix) Its function is to achieve spatial alignment, ensuring that the text vectors Since the latent state vectors of the time-series data reside in the same metric space, the subsequent calculation of the matching degree between the two (i.e., the "second matching degree at the logical level") is mathematically valid. For cases where unstructured semantic data includes image data, the system activates a convolutional neural network (CNN) channel for visual feature extraction. In the industrial scenario of this embodiment, image data is often more intuitive than text; for example, a localized high-temperature bright spot in a thermal imaging image directly corresponds to an overheating fault. The system utilizes a pre-defined deep convolutional network (such as ResNet-50 or EfficientNet) to perform multi-layer convolution and pooling operations on the image, extracting visual features from edge textures to high-level semantics. To transform purely visual features into semantic information usable for reasoning, the system performs a visual-semantic space alignment operation, calculated according to the following image semantic vector generation formula: In the formula, The generated image semantic vector represents the device status information contained in the image;

[0088] This represents the raw industrial image data input, such as an RGB image or a grayscale heatmap;

[0089] This represents the feature extraction function of a convolutional neural network, which includes multiple layers of convolution, activation, and pooling operations, and outputs a high-dimensional feature map.

[0090] This indicates a flattening operation, which converts a multidimensional feature map into a one-dimensional feature vector.

[0091] Represents the visual-semantic alignment matrix, which is a learnable linear transformation matrix used to rotate and scale the basis vectors of the visual feature space to the semantic feature space;

[0092] This indicates a vector normalization operation (such as L2 normalization) to ensure that the generated vectors have uniform magnitudes, which facilitates the calculation of cosine similarity.

[0093] Furthermore, visual feature space and semantic feature space are inherently heterogeneous. For example, in visual space, the high pixel similarity between two images may simply be due to similar lighting conditions, not necessarily the same device state; while in semantic space, what we need is similarity in "meaning".

[0094] In the formula The matrix performs this transformation function, learning how to translate visual features into semantic concepts through pre-training on large-scale image-text pair datasets.

[0095] Normalization further enhances the system's robustness by eliminating scale differences in images caused by shooting distance and lighting intensity. This design allows the system to directly utilize camera footage to assist in fault diagnosis, triggering relevant knowledge graph nodes through image features even in the absence of textual descriptions. Finally, based on the actual data type acquired, the system will... or (Or the vector resulting from the fusion of the two) is defined as the final semantic vector, which together with the aforementioned temporal latent state vector constitutes the dual engine driving the isomorphic projection operation, realizing the deep fusion of physical perception and logical cognition.

[0096] Example 3:

[0097] In this embodiment, obtaining the activation energy of the latent state vector and the semantic vector at each graph node corresponding to the discrete entity based on the isomorphic projection operation based on drift gradient and semantic alignment specifically includes:

[0098] Obtain the weighted Mahalanobis distance between the latent state vector and the static feature vector corresponding to the discrete entity, and define it as the first matching degree at the physical level;

[0099] Obtain the semantic similarity between the semantic vector and the attribute description vector of the discrete entity, and define it as the second matching degree at the logical level;

[0100] Obtain the gradient vector of the potential state vector as it changes over time, obtain the drift gradient as it evolves with the state based on the gradient vector, and project the drift gradient onto the main feature direction of the discrete entity to obtain the third matching degree at the trend level.

[0101] The first matching degree, the second matching degree, and the third matching degree are fused to obtain the activation energy of each discrete entity.

[0102] In this embodiment, the process of obtaining the activation energy of each discrete entity further includes the following steps:

[0103] The anomaly score of the current potential state vector is evaluated based on a preset discriminator;

[0104] A reverse inhibition mechanism is constructed, which suppresses the activation energy of all discrete entities belonging to the fault type based on the anomaly score when the anomaly score is within a preset normal distribution range.

[0105] When the anomaly score is not within the preset normal distribution range, the inhibition is lifted and the activation energy of discrete entities belonging to the fault type is enhanced.

[0106] It should be noted that the isomorphic projection operation in this embodiment is essentially to establish a dynamic energy mapping relationship between the continuous numerical space and the discrete symbolic space. In actual industrial application scenarios, relying solely on single-dimensional similarity matching is often insufficient to cope with complex working conditions. For example, simple numerical proximity may ignore semantic contradictions (such as high vibration values ​​but log records as "shutdown for maintenance"), while simple semantic matching lacks the support of real-time data. Therefore, this embodiment designs an isomorphic projection algorithm that integrates physical, logical, and trend dimensions, and introduces an anomaly focusing mechanism based on a discriminator.

[0107] Specifically, the processor first calculates the first matching degree at the physical level, which aims to measure the degree of agreement between the current operating state of the equipment and the standard state represented by each discrete entity in the knowledge graph in terms of numerical distribution. For each discrete entity in the graph (such as the fault mode "outer ring wear of the spindle bearing"), the system pre-stores its static feature vector in the manifold space. The processor calculates the weighted Mahalanobis distance between the potential state vector and the static feature vector. The reason for choosing Mahalanobis distance instead of Euclidean distance is that there are often strong coupling relationships between the various state variables of industrial equipment. Mahalanobis distance can use the covariance matrix of historical data to eliminate the dimensional differences between variables and decouple this correlation, thereby more realistically reflecting the statistical distance between the state point and the distribution center.

[0108] Simultaneously, the processor calculates a second matching degree at the logical level, aiming to verify whether the current unstructured semantic description is consistent with the attribute descriptions of entities in the graph. Each entity in the graph is associated with a set of textual attribute description vectors (such as "high temperature", "whistling", "crack"). The system calculates the semantic similarity (usually cosine similarity) between the aforementioned generated semantic vectors and these attribute description vectors to determine whether the on-site text logs or image features logically support the activation of the entity.

[0109] Furthermore, to capture the gradual evolution characteristics common in industrial faults, the processor further calculates the third matching degree at the trend level. Industrial faults are often not sudden, but rather undergo a process of quantitative change leading to qualitative change. In the early stages of a fault, the state vector may not have deviated too far from the normal range (i.e., the first matching degree is not high), but its rate of change and direction have already shown a deteriorating trend. To address this, the system calculates the gradient vector of the potential state vector over time, i.e., the drift gradient, which characterizes the velocity and direction of the equipment state in the manifold space. Subsequently, this drift gradient is projected onto the principal feature direction of the discrete entity. The principal feature direction refers to the tangent direction of the typical path of state vector evolution during the occurrence of a fault mode. If the current drift gradient highly coincides with the principal feature direction of a fault, it indicates that the equipment is accelerating towards that fault, and the matching degree in this dimension will increase significantly.

[0110] Based on the calculations of the above three dimensions, the system performs a multidimensional fusion operation to generate the basic activation energy for each discrete entity. This process is carried out in accordance with the aforementioned isomorphic projection activation energy calculation formula (Expression 2).

[0111] However, in industrial environments, sensors are often affected by random noise, causing the latent state vector to occasionally deviate from the normal range. To avoid false alarms caused by such non-fault fluctuations, the system further introduces an anomaly focusing mechanism based on a discriminator. The processor uses a preset discriminator (usually trained as part of a generative adversarial network) to evaluate the "authenticity" or "normality" of the current latent state vector. If the discriminator determines that the current data, although fluctuating, still conforms to the data distribution pattern of normal operating conditions (i.e., within the "normal distribution range"), the system will activate a reverse suppression mechanism to automatically reduce the activation energy of all fault-type entities; conversely, if the discriminator determines that the current data distribution is extremely unusual (i.e., "not within the normal distribution range"), the system will de-suppress and increase the weight of faulty nodes. This dynamic adjustment process is achieved through the following final activation energy adjustment formula: In the formula, This indicates the final activation energy after anomalous focusing adjustment, which is directly used to drive subsequent Bayesian graph network inference;

[0112] This represents the basic activation energy obtained from the aforementioned calculation;

[0113] This represents the degree of anomalousness of the current potential state vector output by the preset discriminator. It typically takes values ​​in the range of [0,1], with higher scores indicating greater anomalousness.

[0114] τ represents the preset threshold for the normal distribution range, used to distinguish the boundary between normal fluctuations and abnormal drifts;

[0115] This function represents the sign function, returning 1 (enhancement) when the input is greater than 0 and -1 (suppression) when the input is less than 0.

[0116] This represents the adjustment intensity coefficient, used to control the magnitude of suppression or enhancement.

[0117] In industrial settings, equipment operates normally most of the time. Even minor fluctuations causing slight increases can be mitigated by the system actively reducing these false activations by returning negative values, thus maintaining system silence and significantly lowering the false alarm rate. However, once the anomaly score exceeds the threshold τ, the sign function flips to positive. The system not only retains the base energy but also amplifies the activation value of the faulty node based on the severity of the anomaly (the size of the difference). This design simulates the cognitive process of human experts: under normal conditions, small fluctuations are relatively ignored, but once an anomaly is detected, it is addressed with focused attention. This mechanism allows the system to automatically and smoothly switch between noise-resistant and agile modes, significantly improving diagnostic efficiency under complex operating conditions.

[0118] Example 4:

[0119] In this embodiment, the step of running a dynamic Bayesian graph neural network for probability propagation along the causal relationships of discrete entities on the industrial knowledge graph, with the activation energy as the initial state, specifically includes:

[0120] A fusion propagation logic comprising data-driven and knowledge-driven terms is constructed. The data-driven terms are determined based on the activation energy ratio of each discrete entity and are used to reflect real-time observation facts of the physical world. The knowledge-driven terms are determined based on the weights of predefined causal relationships and semantic attention coefficients between discrete entities in the industrial knowledge graph and are used to reflect expert experience logic.

[0121] An adaptive balancing factor is set to adjust the weight ratio of the data-driven item and the knowledge-driven item according to the quality or confidence of the current multimodal data, so as to obtain the posterior propagation probability.

[0122] The paths in the industrial knowledge graph are optimized based on the posterior propagation probability. The path with the highest cumulative posterior propagation probability is determined as the evolution path, and the endpoint node of the evolution path is determined as the target fault discrete entity.

[0123] It is understandable that, in order to achieve the intelligent switching in Expression 3 above, this invention needs to accurately calculate the adaptive balance factor. The calculation of this factor depends on the assessment of the current multimodal data quality, and is specifically achieved through the following adaptive balance factor calculation formula: In the formula, This represents the sensitivity adjustment coefficient, used to control the response rate of the balance factor as data quality changes;

[0124] M represents the total number of dimensions of the potential state vectors;

[0125] The logarithmic variance of the m-th dimension in the latent state vector output by the encoder is directly derived from the variational inference results in the previous embodiment and is used to quantify the uncertainty of the data.

[0126] This represents the preset confidence baseline threshold, which indicates the system's average expected level of data quality.

[0127] This represents an exponential function, used here to construct a smooth S-shaped switching curve.

[0128] When the industrial environment is harsh and extremely noisy, the preceding variational encoder automatically predicts a large logarithmic variance. In this case, the exponential term in the formula increases, and the denominator increases accordingly, leading to a decrease in the value of β. This means the system automatically reduces the weight of data-driven terms and relies more on knowledge-driven terms (i.e., 1-β increases). Conversely, when the data waveform is clear and the variance is extremely small, β approaches 1, and the system will primarily rely on real-time data for judgment. This adaptive mechanism endows the model with self-awareness in dynamic environments, enabling it to automatically adjust its inference strategy according to the environmental noise level, significantly improving the environmental adaptability of the diagnostic system. After calculating the posterior propagation probability between nodes in the graph, the system's final goal is to output a clear fault evolution path, which requires finding the path with the highest cumulative probability from the complex graph network. The processor uses dynamic programming to determine the final result based on the following path optimization objective function: In the formula, This represents the final determined optimal evolution path, i.e., the fault propagation chain output by the analysis results;

[0129] This represents the set of all possible connected paths originating from the initial abnormal node;

[0130] Representing a path Adjacent parent-child node pairs;

[0131] This represents the posterior propagation probability after taking the logarithm. The logarithmic operation is introduced to transform the multiplication of probabilities into the summation of logarithmic probabilities, preventing floating-point underflow in long-path inference due to excessively small probability values.

[0132] This represents the parameter variable that maximizes the objective function, i.e., finding the path with the highest score.

[0133] Ultimately, the system will calculate The evolution path is determined, and the endpoint of the path is identified as the target fault discrete entity. In this way, the system not only outputs the conclusion of the fault, but also fully reproduces the evolution process of the fault from the cause to the result, realizing the interpretability analysis of the whole link.

[0134] Example 5:

[0135] In this embodiment, the probability propagation process also includes the following temporal memory fusion step:

[0136] In each iteration of the dynamic Bayesian graph neural network, a temporal convolution operator is introduced to extract residual state features within the historical time window;

[0137] The residual state features are fused with the hidden state of the graph nodes at the current moment so that the probability propagation process is constrained by the historical state evolution trajectory, thereby identifying the evolution path of gradual failure.

[0138] It should be noted that, in order to further improve the system's accuracy in identifying gradual faults in industrial equipment during the probability propagation process using dynamic Bayesian graph neural networks, this embodiment introduces a temporal memory fusion step based on the above-mentioned fusion propagation logic.

[0139] In actual industrial production, faults such as gear pitting, bearing fatigue wear, or minor pipeline leaks often exhibit a significant cumulative effect over time, rather than sudden state jumps. If the inference model relies solely on the activation energy at a single moment for judgment, it is highly susceptible to overlooking subtle early signs of the fault or being misled by transient environmental noise. Therefore, when executing each layer iteration update of the dynamic Bayesian graph neural network, the processor does not view each time slice in isolation. Instead, it introduces a temporal convolution operator to extract residual state features within the historical time window and deeply fuses these features with the hidden state of the graph nodes at the current moment.

[0140] Specifically, when the network calculates the latest state of a graph node (i.e., a discrete entity of child nodes) at time t, the system backtracks the node's historical state sequence over a continuous time window. The processor uses a pre-defined one-dimensional convolution kernel to perform a sliding convolution operation on this sequence in the time dimension to extract implicit temporal evolution patterns (such as monotonically increasing trends or periodic fluctuations). This information extracted from history is defined as a residual state feature. Subsequently, the system uses this residual feature as a constraint term, superimposed on the current probability state propagated from the parent node, thereby generating the node's final hidden state. This temporal memory fusion process strictly follows the following node state temporal update formula for calculation:

[0141] In the formula, This represents the hidden state of the graph node y after the update at time t. This state integrates both spatial causal propagation and temporal historical evolution information.

[0142] This represents a nonlinear activation function (usually ReLU or Leaky ReLU), used to increase the nonlinear expressiveness of the model and ensure that the state values ​​are within the effective range;

[0143] y represents the current discrete child node, which is the target node to be updated in this iteration;

[0144] x represents a discrete entity of the parent node in the industrial knowledge graph that has a direct causal connection with the child node y.

[0145] This represents the set of all predecessor parent nodes of child node y;

[0146] This represents the posterior propagation probability from parent node x to child node y, and this value is directly derived from the calculation result of the fusion causal propagation equation in the previous embodiment;

[0147] This represents the graph space transformation weight matrix, used to perform linear transformations on the spatial features passed in from the parent node.

[0148] This represents the hidden state of parent node x at time t;

[0149] This represents the temporal feature fusion weight matrix, used to adjust the degree of influence of historical residual features on the current state;

[0150] This refers to a temporal convolution operator, i.e. a pre-trained one-dimensional convolution kernel, used to capture specific temporal evolution patterns (such as linear growth or exponential bursts).

[0151] * indicates the convolution operator;

[0152] This represents the sequence of historical hidden states of child node y within the time window prior to the current moment, i.e., the source data of the state residual features. Based on the above expression, it has strong adaptability in special scenarios involving intermittent sensor failures or early, weak faults. Specifically, for example, when monitoring the exhaust temperature of a large compressor, if the temperature sensor reading suddenly drops (noise) within a few seconds due to poor contact, relying solely on real-time data for reasoning might misjudge that the fault has disappeared. However, in this solution, because... If there are persistent high-temperature residual features within the extracted historical window (e.g., the past 60 seconds), the convolution operation will output a large positive value. This inertia will forcibly increase the current value. This compensates for the temporary lack of real-time data, thus maintaining continuous detection of overheating faults. Conversely, if it is a normal signal fluctuation, the convolution output value is small due to the lack of a long-term evolution trend, and will not mislead the results. In this way, the probability propagation process is strongly constrained by the historical state evolution trajectory, enabling the system to effectively filter out transient noise and accurately identify evolution paths with gradual characteristics, such as wear and carbon buildup, significantly improving the stability and reliability of complex fault diagnosis in industrial settings.

[0153] Example 6

[0154] The method also includes a closed-loop self-optimization step:

[0155] Obtain user feedback on the output analysis results, including confirmation or correction of the target fault discrete entity;

[0156] The confirmed time-series data segments and the corresponding correct target fault discrete entities are used to construct high-confidence sample pairs;

[0157] The knowledge contained in the high-confidence sample pairs is compressed and updated into the encoder parameters of the temporal variational manifold space.

[0158] It is understandable that in actual industrial production sites, the operating environment of equipment is dynamic and changing, and new failure modes or data features may gradually drift as the equipment ages. A single pre-trained model is difficult to maintain high accuracy throughout its entire lifecycle. Therefore, this embodiment designs a parameter correction mechanism based on feedback to trace back the diagnostic experience of end users to the feature extraction network at the front end.

[0159] This process begins with obtaining user feedback on the output analysis results. In specific application scenarios, after receiving the diagnostic conclusions and corresponding evolution paths pushed by the system, the operations engineer will verify them in conjunction with on-site disassembly results or endoscopic examination videos. The system provides a feedback entry point through a human-computer interaction interface. If the system diagnosis is correct, the engineer clicks "Confirm"; if the system diagnosis is incorrect, the engineer manually corrects it to the correct fault type from the candidate list in the knowledge graph. This operation generates a clear supervisory signal, namely feedback information, which contains the final decision on the discrete entity of the target fault.

[0160] Next, the processor performs a sample construction operation, building high-confidence sample pairs by pairing the confirmed time-series data segments with the corresponding correct target fault discrete entities. The system then backtracks to the time-series data within the original time window that triggered this diagnosis, marking it as a "standard sample," and using the user-confirmed correct fault entity as the absolute truth label for that data. This pair represents the precise mapping relationship between the physical data form and the logical fault conclusion under the current operating conditions, possessing extremely high confidence and serving as the benchmark for the system's self-correction.

[0161] Subsequently, the system performs the core knowledge compression and update steps, compressing the knowledge contained in the high-confidence sample pairs and updating the encoder parameters in the temporal variational manifold space. Here, "knowledge compression" is mathematically represented by adjusting the network weights of the multimodal encoder through optimization algorithms, so that after the specific temporal data is mapped by the encoder, its position in the manifold space can more accurately fall within the distribution area of ​​the correct fault entity.

[0162] Example 7

[0163] As attached Figure 2 As shown, in this embodiment, a database-driven industrial knowledge analysis system is provided, the system comprising:

[0164] A two-layer space construction module is used to construct a two-layer space architecture for industrial knowledge. The two-layer space architecture includes an industrial knowledge graph and a temporal variational manifold space. The industrial knowledge graph is used to store discrete entities and their causal relationships, and the temporal variational manifold space is used to characterize the distribution patterns of continuous temporal data.

[0165] A multimodal data acquisition module is used to acquire multimodal data from the industrial site. The multimodal data includes time-series data of industrial equipment operation and unstructured semantic data describing the state of industrial equipment. The unstructured semantic data includes text data and / or image data.

[0166] The manifold embedding processing module is used to perform manifold embedding processing on the time series data, mapping the time series data to the time series variational manifold space to obtain a latent state vector containing state deterministic features and uncertain fluctuation features.

[0167] The semantic feature encoding module is used to encode the unstructured semantic data into semantic features, transforming the text data and / or image data into semantic vectors that represent current working condition information.

[0168] The isomorphic projection calculation module is used to obtain the activation energy of the latent state vector and the semantic vector at each graph node corresponding to the discrete entity based on the isomorphic projection operation based on drift gradient and semantic alignment.

[0169] The dynamic reasoning module is used to run a dynamic Bayesian graph neural network to perform probability propagation along the causal relationship of the discrete entities on the industrial knowledge graph, with the activation energy as the initial state, in order to obtain the target fault discrete entity of the current equipment state and the evolution path of the target fault discrete entity.

[0170] The result output module is used to output the target fault discrete entity and the evolution path of the target fault discrete entity as analysis results.

[0171] From an overall architecture perspective, the system in this embodiment is organically composed of seven core functional modules. These modules are connected and interact with each other through a high-speed data bus, together forming a closed-loop intelligent analysis platform.

[0172] First, the system's foundational support layer consists of a two-layer spatial construction module. This module is responsible for initializing and maintaining the system's underlying data architecture, constructing a hybrid knowledge space that integrates discrete symbols and continuous numerical values. Specifically, on the one hand, it utilizes graph database technology (such as Neo4j) to construct an industrial knowledge graph, using discrete entities such as equipment components, fault modes, and maintenance measures as nodes, and logic such as fault propagation mechanisms and compositional relationships as edges, forming a structured knowledge network. On the other hand, it uses high-dimensional vector database technology (such as Milvus) combined with variational autoencoder networks to construct a temporal variational manifold space, used to characterize the probability distribution of continuous time-series data under different operating conditions, thereby providing a unified mathematical benchmark for subsequent cross-modal mapping.

[0173] At the data perception layer, the multimodal data acquisition module acts as the system's senses. Through industrial field bus interfaces or wireless networks, it collects multidimensional time-series data on industrial equipment operation in real time, such as sensor signals for vibration, temperature, and pressure. This data reflects the physical fluctuations of the equipment. Simultaneously, this module also connects to the enterprise's operation and maintenance management system to acquire unstructured semantic data describing the equipment's status, including text logs filled out by maintenance personnel, maintenance work orders, and image data captured by on-site monitoring cameras or thermal imagers. This comprehensive acquisition of multi-source heterogeneous data ensures the integrity of the system's analytical input information.

[0174] Entering the feature processing layer, the system performs standardized mapping of heterogeneous data through two parallel modules. The manifold embedding processing module focuses on the deep analysis of time-series data, and it has built-in pre-trained multimodal encoders and variational inference engines, which can map continuous and noisy time-series monitoring data into the aforementioned time-series variational manifold space. It not only extracts the mean center of the data to determine the state trend, but also calculates the logarithmic variance of the data to quantify the uncertainty brought by environmental noise, and finally generates a latent state vector containing both deterministic state features and uncertainty fluctuation features. At the same time, the semantic feature encoding module is responsible for processing unstructured data. It uses natural language processing models to parse key entities and associations in text, uses convolutional neural networks to extract visual features in images, and encodes this information into a semantic vector with the same dimension as the latent state vector, thereby achieving the initial alignment of physical signals and logical descriptions in the feature space.

[0175] In the fusion inference layer, the isomorphic projection calculation module acts as a bridge connecting the continuous space and the discrete graph. It performs isomorphic projection operations based on drift gradients and semantic alignment, comprehensively calculating the physical drift characteristics of the latent state vectors, the logical consistency of the semantic vectors, and the trend characteristics of state evolution, thereby accurately calculating the activation energy of each discrete entity node in the knowledge graph at the current moment. This mechanism illuminates the previously silent knowledge graph with real-time data, forming a dynamic energy distribution map. Subsequently, the dynamic inference module takes over these activation energies, using them as the initial state, and runs a dynamic Bayesian graph neural network on the causal topology of the industrial knowledge graph. This module integrates data-driven and knowledge-driven adaptive fusion algorithms, dynamically adjusting the inference strategy according to data quality, and combining a temporal memory fusion mechanism to propagate probabilities along the causal relationship chain, ultimately deriving a fault evolution path with the highest cumulative posterior probability and the target fault discrete entity at the path's endpoint.

[0176] Finally, at the application interaction layer, the results output module transforms the complex mathematical results obtained through reasoning into an analysis report that users can understand. It not only outputs the specific target fault name (discrete entity) but also visualizes the fault's evolution path, revealing the complete process from cause to effect. This analysis result is then transmitted to a remote control terminal or on-site handheld devices to assist maintenance personnel in making accurate decisions. Through the close collaboration of these modules, the system successfully achieves end-to-end intelligent analysis from multimodal raw data to interpretable fault conclusions.

[0177] Example 8

[0178] To achieve the above objectives, this embodiment also provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program.

[0179] To achieve the above objectives, this embodiment also provides a computer-readable storage medium storing a computer program, which is executed by a processor.

[0180] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a device including one or any combination of the above-mentioned memories. The computer may be a variety of computing devices, including smart terminals and servers.

[0181] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0182] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0183] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0184] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0185] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0186] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a multimedia terminal device (which may be a mobile phone, computer, television receiver, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0187] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An industrial cognitive analysis method driven by vector and time-series databases, characterized in that, The method includes the following steps: A two-layer spatial architecture for industrial knowledge is constructed, comprising an industrial knowledge graph and a temporal variational manifold space. The industrial knowledge graph is used to store discrete entities and their causal relationships, while the temporal variational manifold space is used to characterize the distribution patterns of continuous temporal data. Acquire multimodal data from the industrial site, including time-series data of industrial equipment operation and unstructured semantic data describing the state of industrial equipment, including text data and / or image data; Manifold embedding processing is performed on the time series data to map the time series data onto the time series variational manifold space to obtain a latent state vector containing state deterministic features and uncertain fluctuation features. Semantic feature encoding is performed on the unstructured semantic data to transform the text data and / or image data into semantic vectors representing current working condition information; The activation energies of the latent state vector and the semantic vector at each graph node corresponding to the discrete entity are obtained through isomorphic projection operations based on drift gradient and semantic alignment. Specifically, this includes: obtaining the weighted Mahalanobis distance between the latent state vector and the static feature vector corresponding to the discrete entity, and defining it as the first matching degree at the physical level; obtaining the semantic similarity between the semantic vector and the attribute description vector of the discrete entity, and defining it as the second matching degree at the logical level; obtaining the gradient vector of the latent state vector over time, obtaining the drift gradient based on the gradient vector as the state evolves, and projecting the drift gradient onto the main feature vector of the discrete entity. The process involves moving upwards to obtain a third matching degree at the trend level; fusing the first matching degree, the second matching degree, and the third matching degree to obtain the activation energy of each discrete entity; and further including the following steps when obtaining the activation energy of each discrete entity: evaluating the anomaly score of the current potential state vector according to a preset discriminator; constructing a reverse suppression mechanism, whereby, when the anomaly score is within a preset normal distribution range, the activation energy of all discrete entities belonging to the fault type is suppressed according to the anomaly score; and when the anomaly score is not within the preset normal distribution range, the suppression is lifted and the activation energy of discrete entities belonging to the fault type is enhanced. Using the activation energy as the initial state, a dynamic Bayesian graph neural network is run along the causal relationship of the discrete entities on the industrial knowledge graph to perform probability propagation in order to obtain the target fault discrete entity of the current equipment state and the evolution path of the target fault discrete entity. The target fault discrete entity and its evolution path are output as analysis results.

2. The industrial cognitive analysis method based on vector and time-series databases as described in claim 1, characterized in that, The step of performing manifold embedding processing on the time-series data, mapping the time-series data to the time-series variational manifold space to obtain a latent state vector containing both deterministic and uncertain fluctuation features, specifically includes: Feature extraction is performed on the time series data within the current time window based on a preset multimodal encoder; The mean vector and variance vector, which characterize the current data distribution, are obtained from the output of the multimodal encoder according to the variational inference mechanism. The latent state vector is obtained by combining reparameterization with Gaussian noise sampling, and then by using the mean vector and variance vector, so that the latent state vector can accommodate random disturbances and noise uncertainties in the industrial environment.

3. The industrial cognitive analysis method based on vector and time-series databases as described in claim 1, characterized in that, The semantic feature encoding of the unstructured semantic data specifically includes: When the unstructured semantic data contains text data, a preset language processing model is used to extract text keywords and their relationships, and these are mapped into a high-dimensional vector as the semantic vector. When the unstructured semantic data includes image data, visual features are extracted from the image data using a convolutional neural network, and these visual features are aligned into the semantic space. The resulting high-dimensional vector is then defined as the semantic vector.

4. The industrial cognitive analysis method based on vector and time-series databases as described in claim 1, characterized in that, The step of using the activation energy as the initial state and running a dynamic Bayesian graph neural network along the causal relationships of the discrete entities on the industrial knowledge graph for probability propagation specifically includes: A fusion propagation logic comprising data-driven and knowledge-driven terms is constructed. The data-driven terms are determined based on the activation energy ratio of each discrete entity and are used to reflect real-time observation facts of the physical world. The knowledge-driven terms are determined based on the weights of predefined causal relationships and semantic attention coefficients between discrete entities in the industrial knowledge graph and are used to reflect expert experience logic. An adaptive balancing factor is set to adjust the weight ratio of the data-driven item and the knowledge-driven item according to the quality or confidence of the current multimodal data, so as to obtain the posterior propagation probability. The paths in the industrial knowledge graph are optimized based on the posterior propagation probability. The path with the highest cumulative posterior propagation probability is determined as the evolution path, and the endpoint node of the evolution path is determined as the target fault discrete entity.

5. The industrial cognitive analysis method based on vector and time-series database as described in claim 1, characterized in that, The method also includes a closed-loop self-optimization step: Obtain user feedback on the output analysis results, including confirmation or correction of the target fault discrete entity; The confirmed time-series data segments and the corresponding correct target fault discrete entities are used to construct high-confidence sample pairs; The knowledge contained in the high-confidence sample pairs is compressed and updated into the encoder parameters of the temporal variational manifold space.

6. An industrial cognitive analysis system driven by vector and time-series databases, characterized in that, The system includes: A two-layer space construction module is used to construct a two-layer space architecture for industrial knowledge. The two-layer space architecture includes an industrial knowledge graph and a temporal variational manifold space. The industrial knowledge graph is used to store discrete entities and their causal relationships, and the temporal variational manifold space is used to characterize the distribution patterns of continuous temporal data. A multimodal data acquisition module is used to acquire multimodal data from the industrial site. The multimodal data includes time-series data of industrial equipment operation and unstructured semantic data describing the state of industrial equipment. The unstructured semantic data includes text data and / or image data. The manifold embedding processing module is used to perform manifold embedding processing on the time series data, mapping the time series data to the time series variational manifold space to obtain a latent state vector containing state deterministic features and uncertain fluctuation features. The semantic feature encoding module is used to encode the unstructured semantic data into semantic features, transforming the text data and / or image data into semantic vectors representing current working condition information. The isomorphic projection calculation module is used to obtain the activation energy of the latent state vector and the semantic vector at each graph node corresponding to the discrete entity based on the isomorphic projection operation based on drift gradient and semantic alignment. Specifically, it includes: obtaining the weighted Mahalanobis distance between the latent state vector and the static feature vector corresponding to the discrete entity, and defining it as the first matching degree at the physical level; obtaining the semantic similarity between the semantic vector and the attribute description vector of the discrete entity, and defining it as the second matching degree at the logical level; obtaining the gradient vector of the latent state vector changing over time, obtaining the drift gradient as the state evolves based on the gradient vector, and projecting the drift gradient onto the discrete entity. In the main feature direction of the entity, a third matching degree at the trend level is obtained; the first matching degree, the second matching degree, and the third matching degree are fused to obtain the activation energy of each discrete entity; when obtaining the activation energy of each discrete entity, the following steps are also included: evaluating the anomalousness score of the current potential state vector according to a preset discriminator; constructing a reverse suppression mechanism, when the anomalousness score is within a preset normal distribution range, suppressing the activation energy of all discrete entities belonging to the fault type according to the anomalousness score; when the anomalousness score is not within the preset normal distribution range, releasing the suppression and enhancing the activation energy of discrete entities belonging to the fault type; The dynamic reasoning module is used to run a dynamic Bayesian graph neural network to perform probability propagation along the causal relationship of the discrete entities on the industrial knowledge graph, with the activation energy as the initial state, in order to obtain the target fault discrete entity of the current equipment state and the evolution path of the target fault discrete entity. The result output module is used to output the target fault discrete entity and the evolution path of the target fault discrete entity as analysis results.

Images