Process device state abnormality detection method for industrial process control

Abstract

The application discloses a process device state anomaly detection method for industrial process control. The method first collects multiple process variables output by process measurement instruments in a production device and aligns the multiple process variables into a multivariate time series window, compresses several measurement point channels into a few spatial cluster nodes, associates each spatial cluster node with a functional unit of the production device, merges the current time window into a few time modal nodes, and then constructs a cluster node graph by combining the two. On the graph, each node simultaneously receives two types of neighbor messages from the spatial direction and the time direction, wherein the connection strength of the spatial edge is recalculated in the current window, and the connection strength of the time edge remains constant in the reasoning stage. The weights of the two types of edges are included in the same normalization formula to compete for the message aggregation budget, and a gating update unit determines the absorption proportion of neighbor information. The number of nodes of the cluster node graph is much smaller than that of the graph built according to the channel and the time stamp, and the overall reasoning overhead is significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a method for detecting abnormal status of process devices in industrial process control. Background Technology

[0002] Process industry production units, such as ethanol distillation units, water treatment systems, and generator sets, typically deploy dozens to hundreds of process measurement instruments. These instruments continuously collect data on changes in process variables such as temperature, pressure, flow rate, and liquid level over time, and integrate this data into a distributed control system (DCS) or programmable logic controller (PLC) for monitoring and regulation. From a process structure perspective, these measurement instruments are arranged according to functional units: measurement points within the same functional unit have strong correlations, while those across functional units have weaker correlations. Meanwhile, the entire production process usually switches between only a few stable operating conditions, such as start-up heating, normal production, load transition, and shutdown. Therefore, anomalies in production units often manifest as collective deviations of several functional units under specific operating conditions, rather than isolated jumps at a single measurement point. Timely detection of such anomalies and locating the relevant functional units is crucial for ensuring safe operation of the unit and preventing product quality fluctuations.

[0003] In recent years, graph neural networks have been increasingly used to process this type of multivariate time series anomaly detection, but existing methods generally have two shortcomings.

[0004] Firstly, there's the scale of the graph. A representative approach treats each sensor channel as a node in the graph, learning dependencies between sensors through attention and identifying anomalies based on the discrepancy between neighbors' predicted values ​​and actual observations. This type of method builds the graph only at the sensor channel level, where the number of nodes equals the number of channels. As the number of channels increases, the inference overhead also increases, and it only models spatial dependencies, without explicitly modeling the operational conditions on the time axis. Another type of method connects the sensor to each timestamp to build a graph, with the number of nodes being... Therefore, the scale of this graph and the cost of reasoning are unbearable in high-channel-count scenarios. Regardless of the type mentioned above, the nodes all fall at the level of the original channel or the channel and timestamp, and there is currently no method to first abstract the channel into a smaller number of cluster nodes before reasoning.

[0005] Secondly, there's the issue of how spatial and temporal dependencies are handled. Some methods set up two branches to calculate attention in the spatial and temporal directions respectively, but these two attention sets are independent, each allocating weights in its normalization process without affecting each other. This means the model cannot dynamically determine the emphasis on spatial relevance and operational continuity based on the actual situation of the current window; the message budgets in both directions are fixed.

[0006] Furthermore, standard graph attention, when updating a node, essentially incorporates a weighted average of neighboring features. This approach has a side effect: when a node exhibits an anomaly or deviates from its normal value, the average of its still-normal neighbors can dilute the anomaly signal, making the anomaly less noticeable and reducing detection sensitivity. Moreover, in the aforementioned all-temporal graph method, the edge weights in the time direction continuously change with the window, lacking a stable characterization of the inherent switching relationships between operating conditions, and the model's perception of operating condition transitions is also unstable. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for detecting abnormal states of process devices in industrial process control.

[0008] The objective of this invention is achieved through the following technical solution: a method for detecting abnormal status of process equipment in industrial process control, comprising: Data from each measuring instrument is collected at a fixed sampling period and aligned into a multivariate time series window. The time series features of each channel within the window are extracted separately. Spatial clustering is performed on the temporal characteristics of each channel to obtain spatial clusters representing the set of associated measurement points, and the spatial clusters are labeled with the functional units of the industrial production device. Temporal modality clustering is performed on the obtained spatial cluster representation, with each temporal modality representing a running state pattern; Arrange the spatial cluster nodes and temporal mode nodes into a mesh graph, so that each node in the mesh graph represents the state in the corresponding spatial cluster and running state mode. The edges in the mesh graph include: spatial edges connecting different spatial cluster nodes in the same running state mode, and temporal edges connecting nodes in different running state modes under the same spatial cluster. Graph reasoning is performed: the spatial cluster representation of the node, the corresponding modal prototype, and the modal context vector of the current window are spliced ​​together and mapped by a multilayer perceptron as the initial embedding of the node; the connection strength of spatial edges and temporal edges are calculated respectively, and then competitive attention of the two types of edges is performed to obtain the attention coefficients competing under a unified denominator. After the neighbor embeddings are weighted and aggregated according to the attention coefficients, they are handed over to the gating update unit to update the node. After iterative updates are performed and the number of updates reaches a preset number of layers, the node is embedded into the input scoring header to obtain the window anomaly score. When the window anomaly score exceeds a preset threshold, the relevant functional unit and operating status mode are determined based on the anomaly evidence matrix, and an alarm signal is output to the industrial control system to prompt the operator to review the corresponding functional unit.

[0009] Furthermore, the specific process of spatial clustering includes: setting a set of learnable spatial prototypes, soft-assigning each channel embedding according to its similarity to each prototype to obtain an assignment matrix, and compressing the N-way measurement points accordingly. The representation of spatial clusters, where each spatial cluster is associated with a functional unit of the production device, here... The values ​​are adaptively selected based on the number of measurement points, and the same allocation matrix is ​​reused during scoring to project abnormal evidence back to a single channel along the spatial allocation relationship.

[0010] Furthermore, the specific process of the temporal modality clustering includes: performing an attention pooling operation on the obtained spatial cluster representation to obtain a summary vector summarizing the overall state of the current window, and then soft-assigning it to... A global modal prototype is used to output the weights of the current window falling on each running state mode. and a modal context vector (This vector is the weighted sum of all global modal prototypes with weights π); number of modes Take the number of pre-set operating state modes for the industrial process.

[0011] Furthermore, the connection strength calculation of the spatial edge is specifically as follows: The representations of the two spatial clusters within the current window are processed by a bilinear matrix. The effect is then mapped to Sigmoid. Interval: in, The strength of the connection of spatial edges, subscript Representing spatial neighbors The corresponding space cluster.

[0012] Furthermore, the connection strength calculation of the time edge is specifically as follows: Modal prototypes fixed after training convergence via bilinear matrix The effect is: in, The connection strength of the time edge, index Representing Time Neighbors The corresponding time mode.

[0013] Furthermore, the competitive attention specifically includes: For any node First, merge its spatial neighbor set and temporal neighbor set into a unified neighbor set. Then, for each neighbor in the set, an attention bilinear matrix is ​​applied. Calculate the compatibility between the current node and its neighbor, and perform spatial neighbor calculations. Neighbors of Time The unnormalized scores are given by the following formulas: The two types of scores are then fed into the same normalized exponential function to obtain competing attention coefficients under a unified denominator, as shown below: in Take the scores from the corresponding neighbors. The attention coefficient allocated to that neighbor.

[0014] Furthermore, the update node specifically refers to: Weighted aggregation is performed on the neighbors to obtain the aggregated message. The node is then updated by the gating update unit, as shown below: The gating update unit is embedded with the current node. As a hidden state, it aggregates messages. As input, the degree to which neighbor context is absorbed is determined by its internal update and reset gates.

[0015] Furthermore, the calculation of the window anomaly score includes: Maintain a reference embedding that tracks the normal state using an exponential moving average, with a decay factor denoted as γ (0 < γ < 1). This reference embedding is denoted as... (The update method is detailed below.) It updates window by window using an exponential moving average and only executes when the current window is judged to be normal. It subtracts the current feature of each node from its corresponding reference embedding and takes the L2 norm to obtain the elements of the anomaly evidence matrix. Then, it uses the modal weights of the current window to update the matrix. Weighted aggregation, followed by normalization based on the number of spatial clusters, yields the scalar window anomaly score, as shown below: in For scale The elements of the abnormal evidence matrix, For the number of spatial clusters, For the number of time modes, For scalar window anomalies, This refers to the updated node features.

[0016] According to another aspect of the specification, a process device state anomaly detection device for industrial process control is also provided, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the aforementioned process device state anomaly detection method for industrial process control.

[0017] According to another aspect of the specification, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, implements the aforementioned method for detecting abnormal states of process devices for industrial process control.

[0018] The beneficial effects of this invention are: The scale of graph reasoning is greatly reduced, and the cost of a single graph propagation is... The magnitude is far lower than that of building the full graph by channel and timestamp in the worst-case fully connected scenario. When the number of channels is large and the window is long, the number of nodes can be reduced from thousands to tens, and the saved product size is considerable, thus significantly improving inference efficiency.

[0019] The two types of edges compete for the budget in the same normalization, enabling the message weights in the spatial and temporal directions to adaptively increase or decrease, and this mechanism does not introduce additional parameters.

[0020] Gated updates protect the deviation signals of anomalous nodes, enabling them to generate higher anomalous scores during the scoring phase even after multiple layers of propagation.

[0021] The time edge weights are globally constant, which not only gives them physical interpretability but also avoids the instability caused by window-by-window recalculation.

[0022] The anomaly determination result outputs an alarm to the industrial control system and locates the corresponding functional unit, enabling operators to conduct targeted inspections and handling of the process equipment in that functional unit, shortening the time from anomaly detection to handling, and preventing the abnormal operating conditions from escalating. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall structure of the method of the present invention; Figure 2 This is a schematic diagram of the dual-type edge structure of the spatiotemporal cluster node graph in this invention; Figure 3 This is a schematic diagram of the dual-type edge competition attention and gating update process of cluster nodes in this invention; Figure 4 This is a flowchart illustrating the reasoning process of spatiotemporal cluster graph reasoning in this invention; Figure 5 This is a schematic diagram of the device that carries out the method in this invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0025] Referring to Figure 1, this embodiment uses an ethanol distillation unit as an example to illustrate the implementation process of this method. The unit consists of functional units such as a feed preheating unit, a column separation unit, a column top condenser / reflux unit, and a column bottom reboiler unit. Process measuring instruments are deployed on it, including temperature sensors installed on each tray, pressure transmitters located at the column top and bottom, flow meters on the feed and reflux lines, a steam flow meter in the reboiler, and a cooling water temperature sensor in the condenser. The outputs of each measuring instrument are collected by a distributed control system (DCS) or programmable logic controller (PLC) at a fixed sampling period and aligned into a multivariate time series window, which serves as the input to this method. Structurally, this method consists of five interconnected parts: data acquisition and encoding, spatial clustering, temporal modal clustering, spatiotemporal cluster graph reasoning, and scoring and alarming. Its input is a multivariate time series window, with the number of channels denoted as . Window length is Feature dimension is .

[0026] For ease of description, a unified convention is first established for the main symbols involved in this embodiment. The same symbols in the formulas, parameters, and embodiments herein will have the same form and meaning. The spatial cluster index is denoted as... The time modality index is denoted as ;No. The representation of a spatial cluster is denoted as , No. The time modal prototype is denoted as The modal context vector of the current window is denoted as The current window is at position 1. The weights on each modality are denoted as follows: Nodes in the cluster node graph In the The embedding of a layer is denoted as Spatial edge weights are denoted as The time edge weight is denoted as ; , , For learnable bilinear matrices, all three are... 3D real matrix The sigmoid function is used; the channel embedding matrix is ​​denoted as... , its first OK For the first The embedded space prototype of the road passage is denoted as The soft allocation matrix from channel to spatial cluster is denoted as Its elements Indicates the first Road access allocated to the first The weights of each spatial cluster, denoted by the soft-assignment temperature coefficient, are given by... The temperature coefficient of time-mode soft assignment is denoted as The spatial cluster representation matrix is ​​denoted as , its first The above spatial cluster representation The number of propagation layers in the graph is denoted as... The reference embedded exponential moving average decay factor is denoted as The threshold for determining window anomalies is denoted as θ; the channel index is denoted as... Values Reference embedding is denoted as The normal-state reference embedding maintained by exponential moving average is a running state variable and not a learnable parameter; throughout the text, the L2 norm of the vector is represented by ||·|, and the superscript is used to denote the L2 norm of the vector. To represent the matrix transpose, with This indicates vector concatenation.

[0027] The aforementioned learnable bilinear matrix, along with other learnable parameters in this method (including the spatial prototype, temporal modality prototype, channel encoder parameters, multilayer perceptron parameters, and scoring head parameters mentioned later), are all jointly optimized as neural network parameters during the training phase: the preset training loss obtained from the forward computation of training samples is weighted and summed with several regularization terms to form the total loss function. The gradient of the total loss function with respect to all learnable parameters is calculated using the backpropagation algorithm, and updated epoch by epoch using mini-batch stochastic gradient descent (such as the Adam optimizer). This process is repeated for multiple training epochs until the total loss function converges. in The preset training loss term, To impose a diversity penalty on prototypes of each time modality, For the remaining optional auxiliary regular expression terms, , These are the corresponding weighting coefficients.

[0028] After training is completed, all learnable parameters (including the aforementioned bilinear matrix, spatial prototype, and temporal modality prototype) remain fixed during the inference phase and are no longer updated with the input window.

[0029] The first step is channel-level temporal coding. Temporal features are extracted individually for each channel within the window, resulting in a channel embedding matrix of size [value missing]. The purpose of processing each channel independently is to preserve as much individual temporal information as possible before compression, thereby providing high-quality input for subsequent clustering.

[0030] The second step is spatial clustering. A set of learnable spatial prototypes is established, and each channel embedding is soft-assigned according to its similarity to each prototype, resulting in an assignment matrix. Based on this matrix, the embeddings are then clustered... Road test points compressed to Representation of a spatial cluster The spatial cluster represents a set of associated measurement points; in this embodiment, the spatial cluster can be exemplarily mapped to a feed preheating unit, a tower separation unit, a tower top condensation reflux unit, and a tower bottom reboiler unit according to a predetermined process unit division. One feasible approach is to adaptively select values ​​based on the number of measurement points. It should be noted that the number of spatial clusters does not need to be equal to the number of functional units. The same functional unit can correspond to one or more spatial clusters, and the specific correspondence is determined by the associated label. (This label can be given in advance by the process personnel based on the process unit division of the device, or automatically determined based on the majority affiliation of the unit to which the member measurement points in each spatial cluster belong.) The same allocation matrix is ​​reused during scoring to project abnormal evidence back to a single channel along the spatial allocation relationship.

[0031] Specifically, the soft allocation is calculated as follows: Embedding of the nth channel within the current window. With the A spatial prototype First, calculate the cosine similarity between the two, then divide by the temperature coefficient. Then, normalize along the spatial cluster dimension to obtain the weight assigned to the spatial cluster by that channel: in, For channel indexing, N is the number of channels; For spatial cluster index, , The number of spatial clusters; Number the spatial clusters used for summation; For the first Embedding vector of the path, For the first Both are spatial prototypes. 3D real vector; is the inner product of the two vectors; ||·| represents the second norm of the vectors. Dividing the denominator by the norms of the two vectors transforms the inner product into a cosine similarity, the value of which falls between -1 and 1. Let be the temperature coefficient, a positive real number. The smaller the value, the closer the assignment is to a hard assignment. To avoid the norm being zero and causing the denominator to be undefined, a very small positive number can be added to the norm during implementation. As can be seen from the above formula, all assignment weights are positive, and the sum of the assignment weights of the same channel to all spatial clusters is always 1.

[0032] Arrange the weights of all channels into a matrix to obtain the aforementioned allocation matrix. , its first Line number Column elements are Arrange the channel embeddings into a channel embedding matrix by row. , its first Act immediately Using the assigned weights as coefficients, the embeddings of each channel are weighted and summed along the channel dimension to obtain the representation of each spatial cluster: in, for transpose, Let be a spatial cluster representation matrix, and its first... Behavior That is, the first in the previous symbol convention Representation of spatial clusters. Since the channel embedding changes with the current window content, the allocation matrix... Representation of spatial clusters Each window is recalculated; spatial prototype The parameters learned during the training phase remain fixed during the inference phase.

[0033] The third step is temporal modality clustering. This involves clustering the results obtained in the previous step... Each spatial cluster represents a single attention pooling operation, yielding a summary vector that encapsulates the overall state of the current window. This vector is then soft-assigned to... A global modal prototype is used to output the weights of the current window falling on each running state mode. and a modal context vector .

[0034] Specifically, let the summary vector obtained from attention pooling be denoted as... The weight of the current window falling on each running state mode is calculated using the following formula: in Temperature coefficient for soft assignment of time modes; j is the time mode number used for summation, traversing all K_T time modes; modal context vector. Take the prototype of each global modality according to the weights The weighted sum, i.e.: Modal number A smaller fixed value, such as 4, is chosen because most industrial processes switch between only a few steady states. For example, in an ethanol distillation unit, start-up heating, normal distillation, feed fluctuation transition, and shutdown can be used as four operating state modes for reference. The time mode is a mode obtained based on data representation and is not limited to a one-to-one correspondence with the above states. It should be noted that the spatial clustering and time mode clustering steps mainly lay the groundwork for subsequent graph reasoning. This invention focuses on protecting the spatiotemporal cluster graph reasoning part.

[0035] The global modality prototype used in the temporal modality clustering, along with the spatial prototype, is a learnable parameter of this method, and their learning processes are consistent: both participate in backpropagation and gradient update together with other parameters such as channel encoder parameters and bilinear matrix during the training phase, without the need for a separate training process; the number of modalities is set according to the number of operating state modes preset in the industrial process, and its value does not depend on a dedicated training step.

[0036] The fourth step is spatiotemporal cluster graph reasoning, which is the core of this embodiment. Referring to Figure 2, the spatiotemporal cluster graph reasoning is performed. spatial cluster nodes and The time modal nodes are arranged according to their combination relationships into a sheet of size [number]. The grid diagram, each node Characterizing the first The spatial cluster in the th ... The state under the various operating state modes. The total number of nodes in the graph and the computational complexity of a single graph propagation are shown in equation (1): in For a cluster of nodes, the number of its elements is the total number of nodes. This represents the time complexity of single-layer message passing; in comparison, the complexity of building a fully connected graph based on channels and timestamps is at most O(n log n). The ratio of the two reflects the extent of the reduction in scale of the present invention.

[0037] The diagram contains two types of edges: solid lines in Figure 2 represent spatial edges, which connect nodes in different spatial clusters under the same operating state mode; dashed lines represent temporal edges, which connect nodes in different operating state modes under the same spatial cluster.

[0038] The initial features of a node are obtained according to equation (2), which is to represent the spatial cluster corresponding to the node. , corresponding modal prototype and the modal context vector of the current window The three signals are spliced ​​together and then mapped by a multi-layer perceptron. dimension: in This represents vector concatenation. For nodes The initial embedding. The three signals respectively carry the state of the subsystem within the current window, the global prior of the running state mode, and the overall running state mode context of the current window. After being concatenated, the initial state of the node contains both spatial and temporal semantics.

[0039] The connection strength of the spatial edges is calculated according to equation (3), which is derived from the representations of the two spatial clusters within the current window via a bilinear matrix. The effect is then mapped to Sigmoid. Interval: Due to the spatial cluster representation The spatial edge weights are recalculated window by window by the encoder. As the window changes, the graph's topology can thus track the fluctuations in the strength of subsystem coupling caused by changes in the operating state mode.

[0040] The connection strength of the time edge is calculated according to equation (4), based solely on the modal prototypes fixed after training convergence. via bilinear matrix The effect is: because The time-bound weights are fixed after training convergence. Maintaining constancy during the inference phase, it characterizes the relatively stable geometric relationships between different operating state modes, such as the degree of proximity between the startup state and the normal load state.

[0041] Then, competitive attention is performed between the two types of edges, the process of which is shown in Figure 3. For any node... First, merge its spatial neighbor set and temporal neighbor set into a unified neighbor set. Then, for each neighbor in the set, an attention bilinear matrix is ​​applied. Calculate the compatibility between the current node and its neighbor. For spatial neighbors... Neighbors of Time The unnormalized scores are given according to formula (5): For simplicity, the subscript of the central node is omitted in equation (5). The left side of both equations refers to the scores between the central node and its corresponding neighbors. Then, the two types of scores are fed into the same normalized exponential function to obtain the attention coefficients of the competition under a unified denominator, as shown in equation (6): Where u and u′ are both neighbor indices in the unified neighbor set Ω. Take the scores of the corresponding neighbors in equation (5), The attention coefficient allocated to the neighbor is given. As can be seen from equation (6), all neighbors in the spatial and temporal directions share the same denominator, so the sum of their budgets is always 1. The model allocates more budget to the direction that is more critical at present, and no additional parameters are introduced in the whole process.

[0042] After obtaining the attention coefficients, the neighbor embeddings are first weighted and aggregated according to the coefficients in equation (6) to obtain the aggregated message. The node is then updated by the gating update unit, as shown in equation (7): The gating update unit is embedded with the current node. As a hidden state, it aggregates messages. As input, the gating system determines the degree to which it absorbs neighbor context through its internal update and reset gates. The key is that when a node becomes abnormal and its current state is significantly inconsistent with its normal neighbors, with proper training, the gating system will tend to retain more of its own state and absorb fewer of its neighbors' states. This abnormal bias is thus protected from being averaged out by normal neighbors during layer-by-layer aggregation. This is crucial for subsequent scoring, because once abnormal nodes are averaged out by their neighbors, they become unrecognizable at the scoring stage.

[0043] The above message passing is repeated. Layer, in this embodiment, is taken as Two layers are sufficient to cover a sufficiently distant neighborhood. The entire inference process can be seen in Figure 4: First, construct the cluster node graph and initialize the nodes, then calculate the spatial edge weights and determine the temporal edge weights in sequence, followed by competitive attention and gating aggregation, and determine whether the preset number of layers has been reached. If not, return to the competitive attention step to continue to the next layer; if the preset number of layers has been reached, output all updated node features and send them to the scoring head.

[0044] Step 5: Scoring and Alarm. After the graph propagation is complete, each node receives an updated set of features. This embodiment maintains a reference embedding that tracks the normal state using an exponential moving average, with a decay factor denoted as γ (0 < γ < 1, taken as 0.99 in this embodiment). This reference embedding is denoted as... It updates window by window using an exponential moving average, and only executes when the current window is determined to be normal. The update formula is as follows: The execution conditions for the above updates differ between the training and inference phases: During training, the training set samples are considered normal, and updates are performed for each window. Alternatively, the moving quantile of the outlier score in the training window can be used as a temporary criterion, and updates are performed only when the score is below that quantile. During inference, a threshold θ defined on the validation set is used as the criterion, and updates are performed only when the outlier score in the current window does not exceed θ, thus avoiding contamination of the normal reference by outlier windows. The elements of the outlier evidence matrix are obtained by subtracting the current feature of each node from its corresponding reference embedding and taking the L2 norm. Then, the modal weights of the current window are used as the basis for further updates. Weighted aggregation and normalization according to the number of spatial clusters yield the scalar window anomaly score, as shown in equation (8): in For scale The elements of the abnormal evidence matrix, This is for scalar window anomalies. By comparing the current window with a threshold θ calibrated on the validation set, it can be determined whether the current window is abnormal. When an abnormality is determined, the elements with the largest values ​​in the abnormality evidence matrix are further selected, and the corresponding functional unit and operating status mode are located according to their row and column indices. Based on this, an alarm signal is output to the distributed control system (DCS) or programmable logic controller (PLC), and the operator station displays which functional unit and operating status mode the abnormality occurred in, prompting the operator to promptly check the corresponding unit. For example, when the evidence matrix points to an abnormality in the top condenser reflux unit under normal distillation operation, the operator is prompted to check the condenser cooling water supply and reflux pipeline status, thereby preventing product purity fluctuations or further development of abnormal operating conditions. Thus, this method translates the analysis results of process data into online monitoring and alarms for the operation of production equipment, forming a complete closed loop from data acquisition, abnormality identification to on-site alarm. In addition, the evidence matrix... It is also possible to project back to a single channel along the spatial allocation relationship to give the contribution of each measuring point to the current anomaly, so as to facilitate subsequent investigation. However, this location analysis is not the focus of this invention.

[0045] The above output results are applied in the field of industrial technology as follows. First, the window anomaly score serves as a real-time quantitative indicator of the operating status of the production unit. It is integrated into the monitoring screen of a distributed control system (DCS) or programmable logic controller (PLC), displayed and archived along with each process variable, allowing operators to monitor the unit's health status in real time. Second, the anomaly evidence matrix provides the distribution of anomalies across two dimensions: functional units and operating status modes. When an alarm occurs, the operator station uses this matrix to display the affected functional units and their operating status modes. Operators can then conduct targeted on-site inspections, repairs, or process parameter verifications of the corresponding process equipment, such as condensers, reboilers, feed lines, valves, pumps, and measuring instruments, eliminating the time loss associated with checking all measuring points one by one. Through these applications, the time from anomaly occurrence to its location and handling is shortened, helping to prevent the expansion of abnormal operating conditions such as product purity deviations, material and energy waste, and even flooding within the tower, thereby improving the operational safety of the production unit and the stability of product quality.

[0046] The following is a typical embodiment to illustrate the effect of this method.

[0047] Regarding data and scenarios, a safety testing platform using an ethanol distillation unit was employed. This platform is equipped with 47 process measurement instruments, covering process variables such as tray temperature, top and bottom pressure, and feed and reflux flow rates. The data incorporates various abnormal operating conditions, including network attacks and load changes. As for parameter settings, the number of spatial clusters was calculated using the method described above. Modal number Feature dimension taken The exponential moving average decay factor is set to γ=0.99, and the number of graph propagation layers is set to... Therefore, the cluster node graph only contains Each node; however, if the entire graph is built based on channels and timestamps, the window length is taken as... The number of time nodes is Compared to the previous method, the node size is reduced by more than a hundred times, and the product required for a single graph propagation is also significantly reduced. In other words, the inference burden is significantly reduced, making it more suitable for real-time operation on field controllers with a large number of measurement points.

[0048] In terms of detection accuracy, measured by the Affiliation-F1 index, our method achieves 0.8818 on this ethanol distillation dataset, which is significantly higher than the corresponding results of graph bias methods and bibranch graph attention methods, ranking first among all the methods compared.

[0049] To clarify the contributions of each core mechanism, ablation comparisons were further performed on the same dataset, and the results are shown in the table below. As shown in the table, replacing the fully connected cluster graph with a sparse graph results in a considerable decrease in accuracy, indicating that in scenarios with many physically coupled subsystems, the fully connected cluster graph makes a significant contribution to accuracy. Similarly, fixing the window-adaptive spatial edges to static values ​​also leads to a decrease in accuracy, demonstrating that window-by-window recalculation of spatial edges does indeed play a role. Both comparisons were conducted independently without altering the gating update and temporal edge mechanisms, thus their conclusions are relatively clear.

[0050] Finally, a supplementary explanation of the generalized implementation is provided. The spatiotemporal cluster graph reasoning part of this invention can be used as a relatively independent graph reasoning component. It only requires the input to be a representation in the form of cluster nodes, that is, the representation of several spatial clusters along with the modality description, without limiting the type of encoder used upstream, nor limiting the downstream task to prediction, classification, or anomaly detection. That is, the two mechanisms of dual-type edge competitive attention and gated neighbor aggregation are decoupled from the upstream and downstream, and can still be built and applied independently after changing the scenario.

[0051] Corresponding to the aforementioned embodiment of a process device state anomaly detection method for industrial process control, the present invention also provides an embodiment of a process device state anomaly detection device for industrial process control.

[0052] See Figure 5 The present invention provides a process device state anomaly detection device for industrial process control, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a process device state anomaly detection method for industrial process control as described in the above embodiment.

[0053] The embodiment of the process device status anomaly detection device for industrial process control provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is used for an abnormal status detection device of a process device for industrial process control provided by the present invention. Except for... Figure 5 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0054] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0055] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0056] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a process device state anomaly detection method for industrial process control as described in the above embodiments.

[0057] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0058] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the aforementioned method for detecting abnormal status of process devices for industrial process control.

[0059] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0060] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for detecting abnormal status of process equipment in industrial process control, characterized in that, include: Data from each measuring instrument is collected at a fixed sampling period and aligned into a multivariate time series window. The time series features of each channel within the window are extracted separately. Spatial clustering is performed on the temporal characteristics of each channel to obtain spatial clusters representing the set of associated measurement points, and the spatial clusters are labeled with the functional units of the industrial production device. Temporal modality clustering is performed on the obtained spatial cluster representation, with each temporal modality representing a running state pattern; Arrange the spatial cluster nodes and temporal modal nodes into a mesh graph, so that each node in the mesh graph represents the state in the corresponding spatial cluster and running state mode. The edges in the mesh graph include: spatial edges connecting different spatial cluster nodes in the same running state mode, and temporal edges connecting nodes in different running state modes under the same spatial cluster. Graph reasoning is performed: the spatial cluster representation of the node, the corresponding modal prototype, and the modal context vector of the current window are spliced ​​together and mapped by a multilayer perceptron as the initial embedding of the node; the connection strength of spatial edges and temporal edges are calculated respectively, and then competitive attention of the two types of edges is performed to obtain the attention coefficients competing under a unified denominator. After the neighbor embeddings are weighted and aggregated according to the attention coefficients, they are handed over to the gating update unit to update the node. After iterative updates are performed and the number of updates reaches a preset number of layers, the node is embedded into the input scoring header to obtain the window anomaly score. When the window anomaly score exceeds a preset threshold, the relevant functional unit and operating status mode are determined according to the anomaly evidence matrix, and an alarm signal is output to the industrial control system to prompt the operator to review the corresponding functional unit.

2. The method for detecting abnormal status of process equipment in industrial process control according to claim 1, characterized in that, The specific process of spatial clustering includes: setting a set of learnable spatial prototypes, soft-assigning each channel embedding according to its similarity to each prototype to obtain an assignment matrix, and compressing the N-way measurement points accordingly. The representation of spatial clusters, where each spatial cluster is associated with a functional unit of the production device, here... The values ​​are adaptively selected based on the number of measurement points, and the same allocation matrix is ​​reused during scoring to project abnormal evidence back to a single channel along the spatial allocation relationship.

3. The method for detecting abnormal status of process equipment in industrial process control according to claim 1, characterized in that, The specific process of the temporal modality clustering includes: performing attention pooling on the obtained spatial cluster representation to obtain a summary vector summarizing the overall state of the current window, and then soft-assigning it to... A global modal prototype is used to output the weights of the current window falling on each running state mode. and a modal context vector (This vector is the weighted sum of all global modal prototypes with weights π); number of modes Take the number of pre-set operating state modes for the industrial process.

4. The method for detecting abnormal status of process equipment in industrial process control according to claim 1, characterized in that, The connection strength calculation for the spatial edge is specifically as follows: The representations of the two spatial clusters within the current window are processed by a bilinear matrix. The effect is then mapped to Sigmoid. Interval: ； in, The strength of the connection of spatial edges, subscript Representing spatial neighbors The corresponding space cluster.

5. The method for detecting abnormal status of process equipment in industrial process control according to claim 1, characterized in that, The connection strength of the time edge is calculated specifically as follows: Modal prototypes fixed after training convergence via bilinear matrix The effect is: ； in, The connection strength of the time edge, index Representing Time Neighbors The corresponding time mode.

6. The method for detecting abnormal status of process equipment in industrial process control according to claim 1, characterized in that, The competitive attention specifically includes: For any node First, merge its spatial neighbor set and temporal neighbor set into a unified neighbor set. Then, for each neighbor in the set, an attention bilinear matrix is ​​applied. Calculate the compatibility between the current node and its neighbor, and perform spatial neighbor calculations. Neighbors in Time The unnormalized scores are given by the following formulas: ； The two types of scores are then fed into the same normalized exponential function to obtain competing attention coefficients under a unified denominator, as shown below: ； in Take the scores from the corresponding neighbors. The attention coefficient allocated to that neighbor.

7. The method for detecting abnormal status of process equipment in industrial process control according to claim 1, characterized in that, The update node is specifically: Weighted aggregation is performed on the neighbors to obtain the aggregated message. The node is then updated by the gating update unit, as shown below: ； The gating update unit is embedded with the current node. As a hidden state, it aggregates messages. As input, the degree to which neighbor context is absorbed is determined by its internal update and reset gates.

8. The method for detecting abnormal status of process equipment in industrial process control according to claim 1, characterized in that, The calculation of the window anomaly score includes: Maintain a reference embedding that tracks the normal state using an exponential moving average, with a decay factor denoted as γ (0 < γ < 1). This reference embedding is denoted as... It updates window by window using an exponential moving average, and only executes when the current window is determined to be normal. The update formula is as follows: ； The elements of the anomaly evidence matrix are obtained by subtracting the current feature of each node from its corresponding reference embedding and taking the L2 norm. Then, the modal weights of the current window are used as the basis for the analysis. Weighted aggregation, followed by normalization according to the number of spatial clusters, yields the scalar window anomaly score, as shown below: ； in For scale The elements of the abnormal evidence matrix, For the number of spatial clusters, For the number of time modes, For scalar window anomalies, This refers to the updated node features.

9. A process device status anomaly detection device for industrial process control, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a process device state anomaly detection method for industrial process control as described in any one of claims 1-8.

10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a process device state anomaly detection method for industrial process control as described in any one of claims 1-8.

Images