A distributed industrial fault diagnosis system and method

Abstract

The application relates to the technical field of fault diagnosis, in particular to a distributed industrial fault diagnosis system and method, which comprises edge terminals and edge servers arranged in various industrial components. The edge terminal is provided with a rapid diagnosis module, a high-level diagnosis module and a dual-mode trigger communication module. The dual-mode trigger communication module independently and in parallel controls uploading of deep feature data according to event triggering and periodic triggering conditions. The edge server is provided with a synchronous sampling module, a feature fusion module and a global diagnosis module, which are used for aligning and fusing multi-source heterogeneous data on a unified time axis, and performing system-level fault diagnosis. The application guarantees real-time response to sudden faults and continuous monitoring of system states through dual-mode triggering, realizes efficient resource utilization and accurate identification of cross-component implicit faults through a hierarchical collaborative architecture, and significantly improves the comprehensive performance of fault diagnosis of complex industrial systems.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically to a distributed industrial fault diagnosis system and method. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In large and complex industrial systems such as process industries and equipment manufacturing, early fault diagnosis and precise fault location are crucial for ensuring safety and preventing unplanned downtime. Currently, data-driven intelligent fault diagnosis methods are gradually replacing traditional threshold alarms and mechanistic model methods, becoming the mainstream technology.

[0004] To address the challenges of widely distributed industrial equipment and massive amounts of data, existing technologies have developed distributed or edge-collaborative diagnostic architectures. These architectures typically deploy lightweight diagnostic models on edge devices close to the data source for local real-time computation, while deploying complex models on the cloud or servers for centralized analysis. However, such distributed diagnostic solutions have inherent flaws in their data collaboration strategies, making it difficult to achieve a good balance between real-time performance, reliability, and systemicity.

[0005] For example, there's the event-triggered collaborative solution. In this solution, the edge terminal only uploads relevant data or preliminary results to the server for in-depth verification when its local model detects a potential fault. While this greatly reduces invalid data transmission and improves real-time response speed, the system's global state awareness relies entirely on the edge terminal's local diagnostic results. When the edge model misses a fault due to noise interference or changes in operating conditions, or when its diagnostic target is an early, weak, or latent fault propagating across components, the relevant data will not be uploaded. This results in the server consistently being unable to obtain the necessary global information flow, thus losing the ability to diagnose cross-component coupled faults and latent faults.

[0006] For example, there's the periodic polling collaborative solution. In this solution, edge terminals ignore their own diagnostic status and upload data to the server at fixed time intervals. This solution ensures that the server continuously receives the global data stream, theoretically enabling system-level analysis. However, this periodic data upload method results in a large amount of normal status data being uploaded to the server, leading to wasted network bandwidth, idle server computing resources, and unnecessary communication delays. When facing sudden, severe faults requiring millisecond-level responses, the fixed upload period introduces unacceptable diagnostic delays, making it difficult to guarantee real-time performance.

[0007] In summary, existing distributed fault diagnosis architectures cannot achieve a good balance between real-time performance, reliability, and system integrity. Event-triggered solutions that focus on real-time performance and efficiency sacrifice system-level diagnostic capabilities, while periodic polling solutions that pursue system integrity sacrifice efficiency and real-time performance. Summary of the Invention

[0008] This invention provides a distributed industrial fault diagnosis system and method, which solves the pain points of centralized fault diagnosis in complex industrial systems, such as poor real-time performance and inability to identify hidden faults across components. It achieves high real-time performance, high reliability, and scalability of system-level fault diagnosis and health management through edge-end collaboration.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The first aspect of the present invention provides a distributed industrial fault diagnosis system, comprising edge terminals deployed on multiple industrial components and at least one edge server;

[0011] Edge terminals include:

[0012] The rapid diagnostic module is used to process the collected component operation data and output preliminary diagnostic results to determine whether there are any abnormalities in the components.

[0013] The advanced diagnostic module extracts deep feature data of components based on runtime data and intermediate outputs from the rapid diagnostic module.

[0014] The dual-mode trigger communication module controls the uploading of deep feature data based on event trigger conditions and periodic trigger conditions. The event trigger condition is "preliminary diagnostic results indicate the presence of an anomaly"; the periodic trigger condition is "reaching the preset global synchronization period". The event trigger conditions and periodic trigger conditions are independent of each other and operate in parallel.

[0015] Edge servers include:

[0016] The synchronous sampling module converts the sampling rate of deep feature data from different edge terminals according to the relationship between the local sampling period of each component and the global synchronization period, and aligns it to a unified global synchronization time axis.

[0017] The feature fusion module fuses the deep feature data of the aligned components to generate system-level fused features;

[0018] The global diagnostic module performs system-level fault diagnosis based on system-level fusion features; system-level faults include latent faults that are transmitted across components.

[0019] Furthermore, during the synchronization sampling module's "alignment to a unified global synchronization timeline" process, it is configured as follows:

[0020] When the When the sampling period of an individual component is less than the global synchronization period, an undersampling method is used, and its calculation method is expressed as follows: ;

[0021] When the When the sampling period of a component is greater than the global synchronization period, an oversampling method based on linear interpolation is adopted, and its calculation method is expressed as follows: ;

[0022] in, Indicates the first The original deep feature sequence of each component Indicates the global synchronization time step The top-aligned first Individual component features, For the first Local sampling period of each component For global synchronization sampling period, For the sampling step index on the global time axis, The symbol represents the consecutive index positions on the component's local time axis that the global time step is mapped to. and These represent the rounding down and rounding up operations, respectively. These are linear interpolation coefficients, used to measure the proportion of contribution of adjacent feature values ​​to the current estimation result.

[0023] Furthermore, before fusion, the feature fusion module is configured to unify the dimensions of the deep feature data of each component to the same dimension through linear mapping if the dimensions of the deep feature data of each component are inconsistent.

[0024] Furthermore, the feature fusion module adopts a weighted fusion method based on an attention mechanism, with its fusion weight coefficients... Through learnable query vectors Component features after unification of dimensions The calculation formula is as follows: Among them, symbols This represents the vector transpose operation.

[0025] Furthermore, the preliminary diagnostic result output by the rapid diagnosis module is an anomaly probability value. In the edge terminal, the adaptive threshold module dynamically calculates and updates the discrimination threshold used for comparison with the anomaly probability values ​​based on the recent historical output anomaly probability value sequence. Its update rules are as follows: ;in, and Let represent the mean and standard deviation of the failure probability within the sliding time window, respectively. This is the sensitivity coefficient.

[0026] Furthermore, the loss function of the advanced diagnostic module during the training phase includes a path consistency loss term. This is used to constrain components that have physical coupling relationships in the system structure. Its feature representation maintains reasonable similarity in the feature space and is defined as follows: ;in, This represents the set of component pairs in a system that have physical or functional coupling relationships. For a set of component indices in the collection, and Each represents a component i With components j High-level feature representation at the same time step The total number of component pairs. This represents the Euclidean second norm.

[0027] Furthermore, the loss function of the advanced diagnostic module during the training phase also includes a temporal smoothing loss term. It is used to suppress non-physical random fluctuations in component features over time, and is defined as follows: ;in, The time window length, For time step index, Indicates components In the Feature representation of each time step.

[0028] Furthermore, the synchronous sampling module is also configured with a fault-tolerant time alignment mechanism: when the component data corresponding to the target global time point t is missing, within the time window... The closest valid sample is selected for compensation or a zero-order preservation strategy is adopted.

[0029] Furthermore, the edge server optimizes and trains the parameters of the global diagnostic module or feature fusion module based on historical diagnostic results, and sends the updated model parameters to the relevant edge terminals to update their local fast diagnostic module or advanced diagnostic module.

[0030] A second aspect of the present invention provides a distributed industrial fault diagnosis method, comprising the following steps:

[0031] Each edge terminal processes component operation data through a rapid diagnostic module and outputs preliminary diagnostic results; each edge terminal extracts deep feature data of the corresponding component through an advanced diagnostic module.

[0032] Each edge terminal determines in parallel whether the event triggering condition and the periodic triggering condition are met;

[0033] When the event triggering condition or the periodic triggering condition is met, the corresponding edge terminal will upload the deep feature data to the edge server.

[0034] The edge server uses a synchronous sampling module to align the received deep feature data of multiple components to a unified global synchronous timeline;

[0035] The edge server uses a feature fusion module to fuse the aligned features and generate system-level fused features.

[0036] The edge server performs system-level fault diagnosis based on system-level fusion features through the global diagnostic module.

[0037] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects:

[0038] 1. By employing a distributed collaborative diagnostic architecture, the core contradiction of traditional methods—the difficulty of simultaneously achieving real-time performance, reliability, and systemicity—is effectively resolved. Through an "independent and parallel" dual-mode triggering strategy and a hierarchical diagnostic model, event-driven millisecond-level rapid response is organically combined with period-driven continuous global state monitoring. This enables the system to provide immediate alerts for sudden faults while simultaneously constructing a complete time-series state flow, thereby reliably capturing early, weak, and latent faults propagating across components. This achieves a balanced and unified diagnostic capability in terms of timeliness and depth.

[0039] 2. By deploying lightweight rapid diagnostic modules and deep feature extraction modules at the edge, the computational burden is effectively shifted forward, significantly optimizing system resource utilization. The rapid diagnostic module can filter out a large number of normal samples locally, uploading only a refined "diagnostic summary" instead of the raw data. This greatly reduces the amount of uplink data, lowers the computational load on the edge server, alleviates network bandwidth pressure, and fully utilizes distributed edge computing power, giving the system excellent scalability and enabling it to flexibly adapt to the addition or removal of device nodes in a modular manner.

[0040] 3. By employing undersampling and interpolation oversampling techniques, the system solves the time alignment problem of multi-source heterogeneous data; combined with fusion methods such as attention mechanisms, it deeply analyzes the coupling relationships between components in a unified feature space. The system can identify system-level composite faults caused by weak abnormal correlations among multiple components, which are difficult to detect using traditional methods, and accurately locate the fault propagation path.

[0041] 4. Adaptive thresholds, a physically constrained loss function, a fault-tolerant alignment mechanism, and an online update process work together to ensure stable system performance under varying noise levels and operating conditions. The system also adapts to equipment aging and process drift through closed-loop learning capabilities, thus maintaining high precision and reliable operation over a long lifespan in complex industrial environments. Attached Figure Description

[0042] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0043] Figure 1 This is a schematic diagram of a distributed industrial fault diagnosis process provided by one or more embodiments of the present invention;

[0044] Figure 2 This is a schematic diagram of the distributed fault diagnosis model architecture provided by one or more embodiments of the present invention;

[0045] Figure 3 This is a schematic diagram of a component network provided in one or more embodiments of the present invention;

[0046] Figure 4 This is a schematic diagram of a global network provided in one or more embodiments of the present invention;

[0047] Figure 5 This is a schematic diagram of a distributed fault diagnosis system architecture provided by one or more embodiments of the present invention;

[0048] Figure 6 This is a schematic diagram of edge reasoning for component X provided in one or more embodiments of the present invention;

[0049] Figure 7 This is a schematic diagram of server inference provided in one or more embodiments of the present invention. Detailed Implementation

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0052] This solution presents a distributed industrial fault diagnosis method that employs a "dual-trigger" collaborative mechanism: event triggering ensures real-time response to sudden faults, while periodic triggering ensures continuous monitoring of system status. By combining component-level rapid screening with global feature fusion, it achieves unified intelligent diagnosis of single-point faults and faults transmitted across components.

[0053] like Figure 1 As shown, a distributed industrial fault diagnosis method includes the following steps:

[0054] The edge terminal collects and preprocesses the operating data of the target industrial component. The preprocessed data is then processed through a lightweight and fast diagnostic network to obtain the preliminary status judgment result of the component. Based on the preliminary status judgment result, a dual-mode triggering strategy is executed to extract the deep feature data of the corresponding component and upload it to the edge server. The status judgment result affects the activation of the periodic triggering channel.

[0055] The edge server receives deep feature data uploaded from one or more edge terminals, generates system-level fused features through time alignment and feature fusion, and performs system-level fault diagnosis analysis based on the system-level fused features.

[0056] The dual-mode triggering strategy is as follows:

[0057] If the initial status assessment result is "abnormal", then the deep feature data of the corresponding component is extracted and uploaded to the edge server;

[0058] And / or, regardless of the initial state assessment result, the deep feature data of the corresponding components are extracted and uploaded to the edge server according to the preset global synchronization cycle.

[0059] Overall, this solution achieves independent diagnosis of each component in the system and captures global latent faults through a layered and collaborative approach. Its core idea is:

[0060] At the component level, lightweight and fast networks are used to filter normal samples, while deep, high-level networks with residual connections are used to achieve fine identification of complex faults.

[0061] At the system level, the high-level features obtained from the diagnosis of each component are unified to the same time scale, and the coupling and transmission relationships between components are captured through the feature fusion module, ultimately locating the source of hidden faults.

[0062] Through this architectural design, the solution can simultaneously achieve efficient local diagnosis and global collaborative analysis, overcoming the difficulties of traditional holistic modeling methods in multi-component, heterogeneous, and cyclical scenarios.

[0063] This solution mainly includes the design of a distributed fault diagnosis model and the design of an edge-to-edge collaborative distributed inference method for the model. The specific solution content is described below.

[0064] The core idea of ​​the distributed fault diagnosis model is to achieve unified identification of single-component faults and cross-component propagated faults in complex systems through a combination of "independent component diagnosis + global collaborative analysis." The overall architecture is as follows: Figure 2 As shown, it includes multiple independent component networks, all of which are connected to the global network.

[0065] The structure of the component network is as follows Figure 3As shown, the system comprises two parts: a fast network and a high-level network, which are interconnected through a residual structure. The fast network is located at the front end of the diagnostic process, receiving preprocessed raw sensor data from the components, extracting basic features, and calculating the failure probability of the current component using a shallow classifier.

[0066] Fast networks take the preprocessed features of the original signals from the components as input, let the input feature vector be... ,in This represents the feature dimension. Fast networks output the probability that the current sample is in an anomalous state through linear discriminant analysis and a nonlinear mapping. Its calculation form is expressed as:

[0067] ;

[0068] in, This represents the probability that the current sample is classified as an abnormal state. The weight vector of the classification layer. For the bias term, the symbol is... This represents the vector transpose operation. This represents a nonlinear normalization function that maps the output to the range of zero to one.

[0069] The system performs rapid screening based on probability values ​​and discrimination thresholds. When the current sample is determined to be in a normal state, the subsequent diagnostic process is terminated; when When a sample is identified as potentially anomaly, its corresponding features are passed to a higher-level network for further analysis. This is the fault detection threshold.

[0070] To improve the model's adaptability and long-term operational stability under different operating conditions, the system introduces a dynamic threshold adjustment mechanism based on a sliding time window. Specifically, within a time window of length... Within a given time window, statistical analysis is performed on the failure probability sequence over a recent period to calculate its mean and standard deviation, and the discrimination threshold is updated accordingly.

[0071] Adaptive threshold at time The update rule is expressed as: ;in, For a moment The corresponding adaptive discrimination threshold, This is the sensitivity coefficient, used to balance the false alarm rate and the false alarm rate. and These represent the mean and standard deviation of the failure probability within the sliding time window, respectively.

[0072] The mean and standard deviation are calculated as follows:

[0073] ;

[0074] in, Indicates the first The failure probability value output by the fast network at each time point. The length of the sliding time window is denoted by , and the summation interval represents the statistical analysis of the probability sequence over a recent period.

[0075] This network acts as a highly efficient front-end filter, quickly filtering out a large number of normal samples, reducing the computational burden on deep networks, and thus significantly improving the overall efficiency and real-time performance of diagnosis.

[0076] Advanced networks perform deep feature extraction on top of fast networks and ensure the effective transmission of shallow information through residual structures. Their input includes not only the deep feature extraction results of the original signal but also the intermediate features output by the fast network. These two are added together using a residual method to form the final component feature representation.

[0077] ;

[0078] in, Indicates the first Each component at time The fusion feature representation, The original input signal, ( ) represents the deep nonlinear feature extraction function. Its model parameters, ( ) represents the residual function that maps intermediate features of a fast network. This is the intermediate feature vector output by the fast network.

[0079] After obtaining the fusion features, the advanced network performs multi-layer nonlinear transformations and classification, outputting the probability distribution corresponding to the specific fault category, which is expressed in the following form:

[0080] ;

[0081] in, Indicates the first The predicted probability for each fault category and These represent the weight matrix and bias term of the classification layer, respectively. The normalization function maps the classification results to a probability distribution. By introducing the residual structure, shallow statistical features and dynamic features can be effectively preserved during deep learning, thereby enabling the model to have stronger feature representation capabilities and convergence stability in complex and multi-type fault scenarios, achieving high-precision component-level fault diagnosis.

[0082] To further enhance the synergistic effect of high-level network output features in system-level diagnosis, this scheme improves the loss function during training by introducing path consistency constraints and temporal smoothness constraints. This makes component-level features more stable and physically consistent in both the temporal and structural dimensions. The improved overall loss function is defined as follows: ;

[0083] in, For the total loss function of the advanced network, For component-level classification loss, For path consistency loss, For time-series smoothing loss, and It is a non-negative balance coefficient used to adjust the relative weight of each loss term in the overall optimization objective.

[0084] The component-level classification loss is used to minimize the error between the model's predicted results and the true fault labels. It is defined as the cross-entropy loss in the following form: ;

[0085] in, Indicates the total number of fault categories. The first in the true label vector The tag value corresponding to the class, the tag uses one-hot encoding. The sample predicted by the model belongs to the first... Output the probability of the type of fault.

[0086] Path consistency loss is used to constrain components that have physical coupling relationships in the system structure. Its feature representations maintain reasonable similarity in the feature space, thereby enhancing the ability to model cross-component associations. It is defined as follows: ;

[0087] in, This represents the set of component pairs in a system that have physical or functional coupling relationships. For a set of component indices in the collection, and These represent the high-level feature representations of component i and component j at the same time step. The total number of component pairs. Describing the Euclidean L2 norm

[0088] Temporal smoothing loss is used to suppress non-physical random fluctuations in component features over time, ensuring that feature evolution conforms to the dynamic laws of equipment operation. It is defined as follows: ;

[0089] in, The time window length, For time step index, Indicates components In the Feature representation of each time step.

[0090] Through the aforementioned joint optimization strategy, path consistency loss can maintain the spatial correlation between features of physically related components when anomalies occur, thus facilitating the global network's capture of cross-component propagating faults; temporal smoothing loss can reduce non-physical jumps in feature changes over time, making feature evolution trajectories more consistent with the actual dynamic characteristics of equipment operation. This improved mechanism effectively enhances the stability, interpretability, and diagnostic reliability of system-level fused features.

[0091] Global network structure such as Figure 4 As shown, it includes a synchronous sampler and a feature fusion module, used to capture transmission faults between different components.

[0092] The synchronous sampling module is used to solve the time alignment problem caused by inconsistent sampling periods of different components. Let the system's sampling period be... The local sampling period for each component is The global synchronization sampling period is The synchronous sampling module maps the feature sequences of each component to a unified map using sampling rate conversion. On the global synchronization timeline, where the time interval is specified.

[0093] When the sampling period of a component is less than the global synchronization period, i.e. At this time, an undersampling method is used. At each global synchronization time step, the latest feature value of the component at the corresponding time point is selected, and its calculation form is expressed as: ;

[0094] in, Indicates the first The original deep feature sequence of each component Indicates the global synchronization time step The top-aligned first Individual component features, This is a discrete-time index on the global synchronization time axis.

[0095] When the sampling period of a component is greater than the global synchronization period, that is... Oversampling is employed. At each global synchronization time point, linear interpolation is performed based on two adjacent feature values ​​of the component to estimate the feature representation at that moment. The calculation form is as follows:

[0096] ;

[0097] in, This indicates the consecutive index positions on the component's local time axis that the global synchronization time step is mapped to. (Symbol) and These represent the rounding down and rounding up operations, respectively. These are the linear interpolation coefficients.

[0098] In real-world industrial scenarios, considering the impact of communication delays, network packet loss, buffer congestion, and clock drift on signal acquisition and transmission, theoretically strict alignment may be difficult to achieve stably. Therefore, this solution proposes a fault-tolerant time alignment mechanism to enhance the robustness and engineering feasibility of global sampling synchronization.

[0099] Specifically, when performing alignment operations, the synchronous sampler first establishes a time window buffer for each component. When data delays or loss are detected in certain components, the system will select the closest time to the target time within the window. Compensation is made using valid samples:

[0100] ;

[0101] in, Indicates the first Each component at the target time The feature representation after compensation, This represents the original feature value of the component at the corresponding timestamp. For the first The set of available timestamps for each component. and These represent the locations closest to and before the target time, respectively. Two valid timestamps, These are the linear interpolation coefficients.

[0102] If continuous packet loss results in no valid samples in the window, the system will adopt a zero-order hold strategy to maintain the feature output from the previous time step and avoid time axis drift.

[0103] Furthermore, to further reduce the cumulative error caused by clock drift, the synchronization sampler is designed with a dynamic time drift correction mechanism. The system periodically detects the clock offset between each terminal and the edge server, and corrects it through a network time protocol or a local clock synchronization module. When the detected offset exceeds a threshold... At that time, the system automatically triggers a resynchronization process to update the time base of each component.

[0104] Through the aforementioned multi-layered protection mechanisms, even in the presence of network fluctuations, signal loss, or time drift, the system can maintain high-precision alignment of multi-component features along a unified time axis, thus providing a reliable foundation for subsequent feature fusion and global fault identification. This design effectively improves the system's stability and real-time performance in complex industrial communication environments, ensuring that the global network can continuously and accurately capture the hidden fault transmission relationships between different components.

[0105] If the feature dimensions of each component are inconsistent, first perform a linear mapping to a unified dimension:

[0106] ;

[0107] in, For the first dimension after mapping to a unified dimension Each component feature vector and These are the weight matrix and bias term corresponding to the linear transformation, respectively. This indicates the th time after mapping to the global timeline. Each component feature vector For the unified feature dimensions, express A dimensional real vector space.

[0108] After completing time alignment and feature dimension unification, the feature fusion module performs cross-component fusion of features from each component. Let the total number of components in the system be... N Then, cross-component feature fusion can be achieved using one of the following two methods.

[0109] The first method is linear weighted fusion.

[0110] Suppose that the characteristic representation of all components satisfies: ;

[0111] The linear weighted fusion feature is calculated as follows: ;

[0112] in, Indicates at time step The system-level fusion feature representation obtained above, For the i-th component, the fixed weight parameters satisfy... The weighting parameters can be determined during the system deployment phase using historical operational data or prior experience.

[0113] Weight It can be determined through training with historical data during the system deployment phase.

[0114] The second approach is weighted fusion based on an attention mechanism. By introducing learnable attention weights, the system can dynamically adjust the contribution levels of different component features according to the current operating conditions. For the... The attention weights for each component feature are calculated as follows:

[0115] ;

[0116] in, For global synchronization time steps Upper Attention weights for each component, satisfying ; These are learnable query vectors used to characterize the global task context; This is the attention mapping matrix.

[0117] Based on attention weights, the system-level fusion feature representation is calculated as follows:

[0118] ;

[0119] in, Indicates at time step The system-level fusion features obtained from the attention mechanism.

[0120] To ensure engineering adaptability, the learning of feature fusion weights and the overall model training adopt a phased training strategy. The training process of the fusion module is consistent with the overall system training process, and the model parameters are gradually updated collaboratively through a phased joint optimization method, specifically including two training phases.

[0121] The first stage is the local pre-training stage. Each component network, including the fast network and the high-level network, is pre-trained separately on a local independent dataset. Only the classification loss is optimized to obtain stable component-level feature representations, thereby ensuring good separability of the feature space.

[0122] The second stage is the global joint training stage. Using the outputs of each component network as input, the synchronous sampler, feature fusion module, and classifier in the global network are activated, and end-to-end joint training is performed based on the total loss function. In this stage, the weights of the fusion layer and the parameters of the global classifier are updated synchronously through backpropagation, enabling the feature fusion module not only to learn the static relationships between components but also to dynamically characterize the interaction relationships between different working conditions using an attention mechanism.

[0123] Finally, the global classifier discriminates the fused features:

[0124] ;

[0125] in, This represents the predicted probability distribution of system-level fault categories. Indicates at time step The system-level fusion features input to the global classifier and These are the weight matrix and bias term of the global classifier, respectively. The normalization function is used to map the classification output into a probability distribution.

[0126] To address the challenges of large-scale, complex industrial systems characterized by numerous devices, wide spatial distribution, and dispersed data sources, this solution proposes a distributed edge-end collaborative reasoning method, employing methods such as... Figure 5 The system architecture shown includes an edge server and multiple edge terminals that communicate with it.

[0127] Edge servers are typically deployed in factory server rooms or engineer workstations, possessing strong computing and storage capabilities; while edge terminals are distributed near various devices or components, generally embedded computing units with basic data processing and model inference functions. This architecture achieves layered collaborative processing of fault detection and diagnosis through "preliminary edge-side diagnosis + centralized edge server inference," effectively reducing the latency and bandwidth pressure caused by centralized data transmission and improving the system's real-time performance and reliability.

[0128] In this architecture, each component's edge terminal deploys a component-level diagnostic model to acquire, preprocess, and quickly screen raw signals; while the edge server deploys a system-level fusion model to receive diagnostic data or synchronization features from multiple edge terminals, further completing the global analysis of complex faults and the identification of faults transmitted across components.

[0129] The edge terminal mainly undertakes local data acquisition, preprocessing, and preliminary inference tasks. The model deployment and inference process on the edge terminal is as follows: Figure 6 As shown, it includes the following steps:

[0130] The sensor collects raw signals during the operation of the component in real time, including information such as vibration, pressure, temperature and current, and generates standardized input data through preprocessing steps such as normalization, filtering and feature compression.

[0131] The preprocessed standardized data is input into the basic feature extraction module to extract basic feature vectors reflecting the equipment's operating status. These features are then fed into the classification layer for rapid binary classification, outputting either "normal" or "faulty".

[0132] When the output result is a normal sample, the inference ends, the data is no longer uploaded, and the system only retains the status record;

[0133] When the result is a faulty sample or the synchronous transmission cycle controlled by the synchronous sampler arrives, the terminal will trigger the residual link module.

[0134] The residual linking module performs differential calculation or residual extraction on the basic features and historical features to generate a residual signal containing information on local state changes. The generated residual data, along with the basic features, is input into the data sending module and sent to the edge server through the communication link according to the set strategy.

[0135] The communication modes are divided into two types:

[0136] Event-triggered mode: Report immediately when the quick diagnostic result indicates a fault;

[0137] Synchronous Trigger Mode: During the synchronous sampling period, residual data from each component are sent uniformly for global analysis.

[0138] In this way, the edge terminal can quickly identify common faults locally, significantly reduce the amount of uplink data, and provide key synchronization feature support for global fault identification.

[0139] The communication modes include event-triggered mode and synchronous-triggered mode, which coexist and represent two different triggering methods for data reporting at the edge. Specifically, they involve two conditional checks; if either condition is met, data is reported. Event-triggered mode means that when the edge terminal detects an anomaly or fault in its local rapid diagnostic module, it immediately and proactively sends diagnostic results and feature data to the edge server, enabling immediate alarms and rapid responses. Synchronous-triggered mode means that each component periodically reports residual features and operating status when a preset synchronous sampling period arrives, used for unified feature alignment across the global network and system-level fault analysis.

[0140] Both communication modes are implemented based on the same communication mechanism and do not require the establishment of independent channels. The only difference lies in the triggering logic: the former is driven by the event detection result, while the latter is triggered by a time scheduling mechanism. This design ensures both real-time reporting of fault events and periodic synchronous data updates required for global diagnosis, thus achieving a balance between communication load and diagnostic timeliness.

[0141] Edge servers are primarily responsible for aggregating data from multiple components, fusing features, and performing global inference; they constitute the system's centralized analysis layer. The model deployment and inference process on the edge server side is as follows: Figure 7 As shown, it includes the following steps:

[0142] After the server receives data from each edge terminal, the data receiving module determines the data source and type:

[0143] If the data is for an event-triggered fault of a single component, the advanced feature extraction module for that component is invoked for deep feature analysis. This module extracts advanced features using a residual-enhanced neural network structure. The corresponding classification layer outputs component-level diagnostic results.

[0144] If the data is batch feature data controlled by a synchronous sampler and involves multiple components, the system sequentially calls the advanced feature extraction module of each component to obtain the aligned advanced feature set. .

[0145] Subsequently, these features are input into the synchronous feature fusion module for multi-component feature fusion.

[0146] During the fusion process, to address the inconsistency in sampling periods and feature dimensions among different components, the system utilizes a synchronized sampler to set a globally unified period. And according to the cycle of each component Oversampling or undersampling is performed to obtain the aligned features, as shown in the following equation:

[0147] ;

[0148] in, This indicates the number of times the global synchronization timeline is reached. Alignment features corresponding to each time step Indicates the first The original deep feature sequence of each component This is a discrete-time index on the globally synchronized time axis. This indicates the consecutive index positions on the component's local time axis that the global synchronization time step is mapped to. (Symbol) and These represent the rounding down and rounding up operations, respectively. These are the linear interpolation coefficients.

[0149] The aligned features are further fused using linear mapping or attention mechanisms, as shown in the following equation:

[0150] , ;

[0151] in, For time steps The overall system characteristics after fusion are represented. For the first The fusion weight of each component, To map to the unified The component feature vector after the feature space is 3D. To synchronize time steps globally The top-aligned first Individual component features, and These are the parameters of the linear transformation.

[0152] This mechanism enables the server to not only identify the mutual influence and potential transmission paths between components, but also to maintain a unified analysis rhythm under the condition of asynchronous acquisition of multi-source data, thereby achieving system-level hidden fault detection.

[0153] This solution achieves significant technical improvements over traditional centralized fault diagnosis methods by constructing a distributed collaborative reasoning architecture on both the end and edge sides, in terms of real-time performance, diagnostic accuracy, cross-component fault identification capabilities, and resource utilization.

[0154] Verification regarding the significant improvement in real-time performance.

[0155] In the Tennessee process simulation verification, the edge-side fast network can complete a fault screening within 30 ms, with a normal sample filtering rate of over 90%. Only abnormal samples and synchronization periodic data are reported to the server, thereby reducing data transmission volume by approximately 80%. The average response latency of the entire edge-side collaborative inference is controlled within 0.5 s, while traditional centralized methods typically require 2–3 s to complete the same diagnosis, improving real-time performance by approximately 4–6 times.

[0156] Validation of improvements in diagnostic accuracy and robustness.

[0157] By introducing a residual enhancement structure and a multi-layer nonlinear feature extraction module, this scheme maintains stable recognition performance under various operating conditions and high noise levels. In tests of 15 typical fault types, including 10 single fault types and 5 composite fault types, the average recognition accuracy at the component level reached 98.2%. The recognition accuracy for system-level hidden transmission faults improved by 15.7% compared to traditional principal component analysis and long short-term memory network centralized models. Within a signal-to-noise ratio range of 10–40 dB, the overall recognition accuracy fluctuated by less than 2%, demonstrating good noise robustness.

[0158] Verification of enhanced cross-component latent fault identification capabilities.

[0159] By achieving multi-source feature time alignment through a synchronous sampler and feature fusion module, this scheme can effectively identify cross-unit fault propagation that traditional models cannot detect. For example, in a simulation scenario, a combined fault of abnormal reactor temperature causing pressure drop in the separation tower is diagnosed by the model of this scheme, with the confidence level increasing from 0.62 in the traditional method to 0.89, thus achieving accurate location of the implicit propagation path.

[0160] Verification of improved resource utilization and system scalability.

[0161] This solution effectively utilizes previously idle edge computing power by deploying a lightweight, high-speed network at the edge, reducing the central server's computing load by approximately 70%. It also allows for flexible addition of component nodes through modular expansion, eliminating the need to retrain the global model. This architecture is suitable for any number of edge terminals and possesses excellent scalability and engineering deployability.

[0162] Verification of the overall effect and evaluation criteria.

[0163] All experiments were conducted using the Tennessee process simulation platform. Evaluation metrics included diagnostic accuracy, processing latency, data upload volume, and bandwidth usage. Compared to representative methods such as centralized LSTM and AE-LSTM, this approach achieved significant improvements in overall performance metrics, with a combined improvement of approximately 2.4 times in real-time performance and diagnostic accuracy.

[0164] Comprehensive experimental results demonstrate that this proposed solution not only effectively overcomes the limitations of traditional centralized methods in terms of insufficient real-time performance and difficulty in identifying latent faults across components, but also significantly improves fault diagnosis accuracy and system operating efficiency. This solution boasts advantages such as reasonable structural design, high operating efficiency, and ease of expansion and deployment, enabling reliable online fault diagnosis and health management in distributed industrial scenarios such as chemical, power, nuclear energy, and intelligent manufacturing.

[0165] The Tennessee process will be used as an example for illustration. The Tennessee process is a typical complex chemical process, which includes multiple functional units such as reactors, separation towers, condensers, reboilers, compressors, and feeding systems. There are significant nonlinear coupling relationships and dynamic correlation characteristics between these units, making it suitable as a standard test platform for verifying the distributed fault diagnosis performance of complex systems.

[0166] To address the distributed nature of the process, this embodiment sets up edge servers and edge terminal nodes at the system architecture level. The edge servers are deployed in engineering workstations or industrial server rooms, possessing strong computing and data storage capabilities. Edge terminal nodes are deployed near each process unit, including reactors, separation towers, compressors, condensation systems, and feeding systems. The edge terminal nodes are used to perform local signal acquisition, preliminary diagnosis, and feature extraction and transmission, thus forming a two-layer inference system combining rapid end-side response and global edge-side collaboration. This system can meet the real-time requirements of industrial sites while also enabling comprehensive identification of complex fault modes.

[0167] In this embodiment, an embedded edge terminal is deployed next to the field control device of each component to execute the component-level real-time diagnostic model. The edge diagnostic model includes a data preprocessing module, a fast feature extraction module, and an advanced feature extraction module.

[0168] The data preprocessing module is used to clean and structure the component's sensor signals in real time. Specifically, it includes:

[0169] Denoising and filtering: A strategy combining sliding window averaging and low-pass filtering is used to remove sampling noise and instantaneous fluctuations;

[0170] Normalization and dynamic standardization: The statistics are updated adaptively based on the historical operating range of the equipment to ensure the stability of the input feature distribution under different operating conditions;

[0171] Temporal window reorganization: A temporal feature matrix is ​​formed by using a fixed time length as a sliding window, providing a unified input dimension for subsequent networks.

[0172] The Fast Network module is a lightweight shallow feature discriminator, mainly composed of two convolutional layers and one fully connected layer. The first convolutional layer extracts temporal local features, the second convolutional layer performs channel compression, and the fully connected layer outputs the anomaly probability.

[0173] Convolutional layer configuration: kernel size is 3, number of channels is 32 and 16 respectively, and LeakyReLU activation function is used;

[0174] Discriminant layer: It adopts a fully connected structure, with the number of nodes mapped from 16 to 1. Through non-linear normalization mapping, the output is restricted to the interval between zero and one, which is used to represent the discrimination result that the current sample belongs to an abnormal state, thereby realizing binary classification diagnosis.

[0175] The output of the fast network not only determines whether the current sample is in a normal operating state, but also generates a set of intermediate feature representations. These intermediate features are fixed-dimensional feature vectors and are passed as residual information to the higher-level network to support subsequent deep feature modeling and joint inference. When the fast network determines that the current sample is in a normal state, the system only outputs the corresponding health status flag without triggering the deep inference process, thereby effectively reducing computational overhead and improving overall diagnostic efficiency.

[0176] The advanced network module primarily undertakes deep diagnostic tasks for complex patterns and consists of multiple convolutional layers, attention modules, and residual connections. Its structural configuration is as follows:

[0177] Input layer: Receives the original temporal features and the intermediate features of the fast network, and matches the dimensions through a linear projection layer;

[0178] Feature extraction layer: A three-layer one-dimensional convolutional structure is adopted, with convolutional kernel sizes of 3, 5, and 7 respectively, and the number of channels of 64, 64, and 128, which is used to extract complex dynamic features across time scales;

[0179] Attention enhancement layer: A lightweight channel attention module is inserted after the convolutional layer to enhance the expression intensity of fault-sensitive signals through a feature weighting mechanism;

[0180] Residual connections: The intermediate features output by the fast network are linearly mapped and added to the features of the high-level network, enabling the transfer of information from shallow layers to deeper layers.

[0181] The last two layers of the advanced network are a fully connected layer and a Softmax layer, used to output the specific fault type and its confidence level for the corresponding component. Simultaneously, this module calculates residual information in real time during inference. The residual information consists of the difference between the current features and the health baseline features, and is transmitted to the global server to support subsequent feature fusion and comprehensive analysis.

[0182] The edge diagnostic module interacts with the edge server through the communication management unit, and the communication uses the lightweight MQTT or industrial OPC UA protocol.

[0183] When a fault is detected or a synchronization period is reached, the system packages the features and residuals into a message frame and sends it.

[0184] To ensure timing consistency, the message frame carries a local timestamp and component number; if the network is interrupted, the edge terminal will automatically cache and retransmit to ensure data continuity.

[0185] In actual simulations of the Tennessee process, the edge model can complete one inference within 30 ms. The fast network filters out an average of 90% of normal data, with only about 10% of samples uploaded to the edge server, significantly reducing communication and computational pressure.

[0186] Edge servers are deployed at factory engineering workstations and have GPU acceleration capabilities to perform global fusion and system-level diagnostics.

[0187] The server-side mainly consists of the following four functional modules: data receiving and synchronization module, feature fusion module, hidden feature extraction module, and global classification module.

[0188] The data receiving and synchronization module receives data packets from multiple terminals and performs synchronous sampling based on timestamps. An asynchronous buffer queue mechanism is employed to handle situations where the sampling periods of different terminals are not synchronized, and linear interpolation is used to align features on the same time axis.

[0189] To address the issue of inconsistent feature dimensions among different components, the feature fusion module first uses a linear mapping layer to uniformly map the features of each component to the same feature dimension. Then, it introduces an attention fusion structure to weightedly integrate cross-component features. The attention weights are automatically learned by the model during training and are used to characterize the degree of influence of each component on the system state at the current moment, thereby achieving adaptive fusion and collaborative representation of multi-source features.

[0190] The hidden feature extraction module, the core of global inference, employs a hybrid structure of temporal convolutional networks and multi-head self-attention. The temporal convolutional network layers extract temporal correlations and dynamic trends; the self-attention layers model global dependencies across components; layer normalization and residual paths are added within the module to ensure gradient stability and training convergence. The final output global feature tensor is compressed through two fully connected layers before being input into the classification module.

[0191] The global classification module outputs system-level fault identification results and confidence levels based on the global representation of hidden features. Output categories include: normal state, single-component fault, compound fault, and cross-component transmitted fault.

[0192] In the Tennessee process simulation verification, each component terminal operated at a sampling rate of 10 Hz, and the server synchronization cycle was 1 s. The overall system operation flow is as follows: each edge terminal independently completes signal acquisition and performs rapid diagnosis in real time; when a potential abnormal sample is detected or the synchronization trigger cycle is reached, the edge terminal sends the corresponding residual information to the server; the server performs time alignment and multi-component feature fusion on the data from different components; subsequently, the hidden feature extraction module performs global inference and outputs system-level fault diagnosis results; finally, the diagnosis results and their confidence levels are sent back to the monitoring system or maintenance decision module. In the experiment, the system achieved an overall identification accuracy of 98.2% for 10 types of single faults and 5 types of composite faults, and significantly improved the identification capability for cross-component transmission faults, including situations such as pressure drop changes in the separation tower caused by abnormal reactor temperature, with an overall identification rate improvement of approximately 15%. At the same time, the average diagnostic latency of the system is controlled within 0.5 seconds, which can meet the real-time requirements of industrial sites.

[0193] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A distributed industrial fault diagnosis system, characterized in that, This includes edge terminals deployed across multiple industrial components and at least one edge server; Edge terminals include: The rapid diagnostic module is used to process the collected component operation data and output preliminary diagnostic results to determine whether there are any abnormalities in the components. The advanced diagnostic module extracts deep feature data of components based on runtime data and intermediate outputs from the rapid diagnostic module. The overall loss function of the advanced diagnostic module during the training phase is a weighted sum of component-level classification loss, path consistency loss, and temporal smoothing loss. The advanced diagnostic module includes a path consistency loss term in its loss function during the training phase. This is used to constrain components that have physical coupling relationships in the system structure. Its feature representation maintains reasonable similarity in the feature space and is defined as follows: ;in, This represents the set of component pairs in a system that have physical or functional coupling relationships. For a set of component indices in the collection, and These represent the high-level feature representations of component i and component j at the same time step. The total number of component pairs. Represents the Euclidean second norm; The loss function of the advanced diagnostic module during the training phase also includes a temporal smoothing loss term. This is used to suppress non-physical random fluctuations in component features over time. Component-level classification loss is used to minimize the error between the model's predicted results and the actual fault labels; The dual-mode trigger communication module controls the uploading of deep feature data based on event trigger conditions and periodic trigger conditions. The event trigger condition is "preliminary diagnostic results indicate an anomaly"; the periodic trigger condition is "reaching the preset global synchronization period". The event trigger conditions and periodic trigger conditions are independent of each other and run in parallel. If the initial status assessment result is "abnormal", then the deep feature data of the corresponding component is extracted and uploaded to the edge server; regardless of the initial status assessment result, the deep feature data of the corresponding component is extracted and uploaded to the edge server according to the preset global synchronization cycle. Edge servers include: The synchronous sampling module converts the sampling rate of deep feature data from different edge terminals according to the relationship between the local sampling period of each component and the global synchronization period, and aligns it to a unified global synchronization time axis. The synchronization sampling module is configured to: During the alignment to a unified global synchronization timeline, the synchronization sampling module is configured to: When the When the sampling period of an individual component is less than the global synchronization period, an undersampling method is used, and its calculation method is expressed as follows: ; When the When the sampling period of a component is greater than the global synchronization period, an oversampling method based on linear interpolation is adopted, and its calculation method is expressed as follows: ; in, Indicates the global synchronization time step The top-aligned first Individual component features, Indicates the first The original deep feature sequence of each component For the first Local sampling period of each component For global synchronization sampling period, This is the index of the sampling step on the global time axis; The symbol represents the consecutive index positions on the component's local time axis that the global time step is mapped to. and These represent the rounding down and rounding up operations, respectively. These are linear interpolation coefficients, used to measure the proportion of contribution of adjacent feature values ​​to the current estimation result; The synchronous sampling module is also configured with a fault-tolerant time alignment mechanism: when the component data corresponding to the target global time point t is missing, within the time window... The closest valid sample is selected for compensation or a zero-order preservation strategy is adopted. The feature fusion module fuses the deep feature data of the aligned components to generate system-level fused features; The global diagnostic module performs system-level fault diagnosis based on system-level fusion features; system-level faults include latent faults that are transmitted across components.

2. The distributed industrial fault diagnosis system as described in claim 1, characterized in that, Before fusion, the feature fusion module is configured to unify the deep feature data of each component to the same dimension through linear mapping if the dimensions of the deep feature data of each component are inconsistent.

3. The distributed industrial fault diagnosis system as described in claim 2, characterized in that, The feature fusion module employs a weighted fusion method based on an attention mechanism, with fusion weight coefficients... Through learnable query vectors Component features after unification of dimensions The calculation formula is as follows: Among them, symbols This represents the vector transpose operation.

4. The distributed industrial fault diagnosis system as described in claim 1, characterized in that, The preliminary diagnostic result output by the rapid diagnostic module is an anomaly probability value. In the edge terminal, the adaptive threshold module dynamically calculates and updates the discrimination threshold used for comparison with the anomaly probability values ​​based on the recent historical output anomaly probability value sequence. Its update rules are as follows: ;in, and Let represent the mean and standard deviation of the failure probability within the sliding time window, respectively. This is the sensitivity coefficient.

5. A distributed industrial fault diagnosis system as described in claim 1, characterized in that, The loss function of the advanced diagnostic module during the training phase also includes a temporal smoothing loss term. Its definition is: ;in, The time window length, For time step index, Indicates components In the Feature representation of each time step.

6. A distributed industrial fault diagnosis system as described in claim 1, characterized in that, The edge server optimizes and trains the parameters of the global diagnosis module or feature fusion module based on historical diagnosis results, and sends the updated model parameters to the relevant edge terminals to update their local fast diagnosis module or advanced diagnosis module.

7. A distributed industrial fault diagnosis method, applied to the distributed industrial fault diagnosis system as described in any one of claims 1-6, characterized in that, Includes the following steps: Each edge terminal processes component operation data through a rapid diagnostic module and outputs preliminary diagnostic results; each edge terminal extracts deep feature data of the corresponding component through an advanced diagnostic module. Each edge terminal determines whether the event triggering conditions and periodic triggering conditions are met in parallel. When the event triggering condition or the periodic triggering condition is met, the corresponding edge terminal will upload the deep feature data to the edge server. The edge server uses a synchronous sampling module to align the received deep feature data of multiple components to a unified global synchronous timeline; The edge server uses a feature fusion module to fuse the aligned features and generate system-level fused features. The edge server performs system-level fault diagnosis based on system-level fusion features through the global diagnostic module.

Images