Controller production equipment state early warning method and system

Abstract

The application provides a controller production equipment state early warning method and system, and relates to the technical field of equipment state monitoring. The method comprises the following steps: acquiring multi-modal time series data through an edge perception node, constructing a topology structure code based on a process flow dependency to obtain a state representation vector, performing probability distribution modeling on the representation vector to generate state prior and posterior distributions, calculating a divergence value as an abnormal confidence label to mark an abnormal state, identifying associated nodes along the topology structure through causal diagram inference and performing counterfactual reasoning to calculate causal effect strength to identify a root cause abnormal source, constructing an early warning propagation path based on the root cause abnormal source along the topology structure, and generating a hierarchical early warning instruction according to the causal effect strength.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring technology, and in particular to a method and system for early warning of the condition of production equipment using a controller. Background Technology

[0002] Equipment status early warning is a key technical aspect of ensuring the continuous and stable operation of a production line. Conventional approaches typically rely on single-sensor threshold monitoring or simple statistical models, such as setting fixed upper and lower limits for physical quantities like temperature and vibration, triggering an alarm when these limits are exceeded. Some solutions employ machine learning models, such as support vector machines or isolated forests, to detect anomalies in historical data, extracting frequency or time domain indicators through feature engineering for discrimination. Simultaneously, some systems incorporate expert rule bases, manually associating typical fault modes with corresponding parameter changes to form the judgment logic.

[0003] However, complex and dynamic process dependencies exist between production equipment. Existing technologies often treat each piece of equipment as an independent entity, ignoring the transmission effect of upstream equipment status changes on downstream equipment. For example, tool wear in a preceding process can gradually lead to deviations in the machining accuracy of subsequent processes. However, relying solely on single-equipment thresholds may fail to provide early warnings, and by the time significant exceedances occur downstream, the fault has already spread to multiple nodes. Isolated analysis leads to delayed warnings and generates numerous false alarms or missed alarms. Furthermore, even when an anomaly is detected, existing technologies struggle to automatically pinpoint the root cause. When multiple sensors show fluctuations simultaneously, maintenance personnel must rely on experience to troubleshoot item by item, lacking a systematic approach to tracing the source of the anomaly from a temporal and causal perspective. Most solutions only output a list of equipment-level alarms, failing to distinguish between causal and correlational relationships. For instance, a temperature increase in a component might be due to load changes, but existing monitoring systems would directly classify it as a fault point, resulting in inaccurate root cause identification. Summary of the Invention

[0004] This invention provides a method and system for early warning of the status of production equipment using a controller, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of the present invention provides a method for early warning of the status of a controller-manufactured equipment, comprising: Multimodal time-series data of production equipment operation is acquired through edge sensing nodes, equipment topology is constructed according to the process flow dependency relationship between production equipment, and equipment state representation vector is obtained by encoding the multimodal time-series data based on the equipment topology. The device state representation vector is modeled with a probability distribution to generate a state prior distribution and a state posterior distribution. The divergence value between the state prior distribution and the state posterior distribution is calculated as the anomaly confidence. Device state representation vectors whose anomaly confidence exceeds the adaptive threshold are marked as anomaly state representation vectors. Using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, the causal graph inference algorithm is used to identify related nodes with causal dependencies along the device topology. The historical device state representation vectors of the related nodes are extracted, and counterfactual reasoning is performed on the historical device state representation vectors to generate comparative state representation vectors. The causal effect strength between the historical device state representation vectors and the comparative state representation vectors is calculated to identify the root cause of the anomaly. Based on the root cause anomaly source, an early warning propagation path is constructed along the device topology. According to the causal effect intensity of each node in the early warning propagation path, a graded early warning command is generated and sent to the device control layer for execution.

[0006] In one alternative implementation, Multimodal time-series data of production equipment operation is acquired through edge sensing nodes. An equipment topology is constructed based on the process dependencies between production equipment. The multimodal time-series data is then encoded based on the equipment topology to obtain an equipment state representation vector, including: Operational signals are collected by edge sensing nodes deployed at various production equipment. The operational signals are filtered and preprocessed, and the multimodal time series data is obtained synchronously according to the timestamp. The multimodal time series data is decomposed into time and frequency domain features. The multimodal time series data is statistically analyzed by sliding window to obtain trend features, which are then concatenated with the frequency domain features to obtain a multimodal feature vector. Extract the node identifiers and edge connections of each production device from the device topology, construct an adjacency matrix based on the edge connections, calculate the degree matrix of the adjacency matrix and normalize it to obtain the propagation matrix, and arrange the multimodal feature vectors according to the node identifiers to form a node feature matrix; The node feature matrix and the propagation matrix are multiplied to obtain the first-order neighborhood features. The first-order neighborhood features are nonlinearly activated and multiplied with the propagation matrix to obtain the second-order neighborhood features. The node feature matrix is ​​then weighted by residuals and linearly projected based on the second-order neighborhood features to obtain the device state characterization vector.

[0007] In one alternative implementation, The device state representation vector is modeled using probability distribution to generate a prior state distribution and a posterior state distribution. The divergence value between the prior state distribution and the posterior state distribution is calculated as the anomaly confidence score. Device state representation vectors whose anomaly confidence scores exceed an adaptive threshold are marked as anomaly state representation vectors. The device state representation vector is nonlinearly mapped to obtain a latent space representation. Based on the latent space representation, a distribution mean and distribution variance are generated through dual-path encoding. A state posterior distribution is constructed based on the distribution mean and distribution variance. After performing a log-space transformation on the distribution variance, a noise vector is sampled from the standard normal distribution and a posterior sampling vector is constructed through reparameterization. The posterior sampling vector is reconstructed through multi-level decoding to obtain a reconstructed representation vector. The reconstruction loss between the device state representation vector and the reconstructed representation vector is calculated and a state prior distribution is constructed through maximum likelihood estimation. The divergence between the posterior and prior state distributions is calculated as the anomaly confidence level. Quantile statistics are performed on the anomaly confidence levels of each device node within a pre-stored historical sliding time window to obtain distribution feature parameters. A dynamic adjustment factor is calculated based on the distribution feature parameters and the current device load state. An adaptive threshold is obtained by adjusting a preset baseline threshold using the dynamic adjustment factor. The anomaly confidence level is compared with the adaptive threshold. Device state representation vectors whose anomaly confidence levels exceed the adaptive threshold are marked as anomaly state representation vectors.

[0008] In one alternative implementation, Using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, the causal graph inference algorithm is used to identify associated nodes with causal dependencies along the device topology, and the historical device state representation vectors of the associated nodes are extracted, including: Node identifiers are extracted from the abnormal state representation vectors. The incoming edge connections of the node identifiers are extracted from the device topology and a candidate predecessor node set is determined. The historical device state representation vectors of each node in the candidate predecessor node set are extracted and the temporal Granger causality test statistic is calculated. Nodes with causal precedence relationships are selected based on the Granger causality test statistic to obtain a first-order causal predecessor node set. For each node in the first-order causal predecessor node set, a predecessor node extraction operation is recursively performed. In each recursion, the incoming edge connection relationship of the current node is extracted to obtain a recursive candidate predecessor node set and the corresponding historical device state representation vector is determined. Based on the historical device state representation vector, the time-series Granger causality test statistic is calculated and indirect causal relationships are eliminated through conditional independence test to obtain multiple direct causal predecessor nodes and construct a multi-order causal predecessor node set. When the recursion depth reaches a preset number of layers, the recursion is terminated and the first-order causal predecessor node set and the multi-order causal predecessor node set are merged to obtain an associated node set. Using the Granger causality test statistic as edge weights, a directed weighted causal graph is constructed between the set of associated nodes and the abnormal nodes corresponding to the abnormal state representation vector. Based on the directed weighted causal graph, the historical device state representation vector of each node in the set of associated nodes is extracted.

[0009] In one alternative implementation, Performing counterfactual reasoning on the historical equipment state representation vector to generate a contrasting state representation vector, and calculating the causal effect strength between the historical equipment state representation vector and the contrasting state representation vector to identify root cause anomalies includes: The causal paths between each associated node and the abnormal node are extracted from the directed weighted causal graph, and the causal strength of the path is calculated based on the edge weight. The associated nodes whose causal strength of the path exceeds a preset threshold are selected as the intervention candidate node set. The device state representation vector corresponding to the intervention candidate node set is extracted from the historical device state representation vector as the original state representation vector. The original state representation vector is variationally encoded and mapped to the latent space to obtain latent distribution parameters. Abnormal latent dimensions in the latent distribution parameters are identified, truncated, and resampled to obtain a corrected latent vector. The corrected latent vector is decoded, mapped, and reconstructed to the original representation space to obtain an intervention state representation vector. The intervention state representation vector is used as the initial input and propagated along the causal path using graph convolution. At each propagation node, the input representation vector is weighted and aggregated according to the edge weights, and the node state representation is updated. The propagation is then performed to the abnormal node position to obtain a comparison state representation vector. The representation space distance between the abnormal state representation vector and the comparison state representation vector is calculated as the counterfactual bias. The counterfactual bias and the path causal intensity of the corresponding node are weighted and normalized to obtain the causal effect intensity. The node with the largest causal effect intensity is selected as the root cause anomaly source.

[0010] In one alternative implementation, The process involves identifying anomalous latent dimensions in the latent distribution parameters, truncating and resampling them to obtain a corrected latent vector, and then decoding and mapping the corrected latent vector back to the original representation space to obtain the intervention state representation vector, including: Extract the distribution mean and distribution variance of each potential dimension from the potential distribution parameters, calculate the deviation of the distribution variance of the potential dimension from the distribution variance of the corresponding dimension in the pre-trained normal state potential space, mark the potential dimensions with the deviation exceeding the preset deviation threshold as abnormal potential dimensions, and extract the corresponding distribution mean and distribution variance to construct an abnormal dimension distribution set. Extract the mean and variance of the normal distribution for each position of the abnormal dimension in the abnormal dimension distribution set from the pre-trained normal state latent space. Construct a normal dimension distribution based on the mean and variance of the normal distribution and perform random sampling to obtain normal dimension sample values. Replace the distribution parameters of the abnormal latent dimension in the latent distribution parameters with the normal dimension sample values ​​to obtain the corrected distribution parameters. Based on the modified distribution parameters, a modified latent vector is constructed and mapped to a high-dimensional feature space through multi-layer nonlinear decoding to obtain intermediate reconstructed features. The intermediate reconstructed features are then dimensionally expanded and mapped to the representation dimension corresponding to the original state representation vector through a linear projection layer to obtain the intervention state representation vector.

[0011] In one alternative implementation, Based on the root cause anomaly source, an early warning propagation path is constructed along the device topology. A tiered early warning command is generated according to the causal effect intensity of each node in the early warning propagation path and sent to the device control layer for execution, including: The outgoing edge connections of the root cause anomaly source are extracted from the device topology to obtain a set of successor nodes. The process flow dependency strength between each successor node and the root cause anomaly source is extracted. Based on the process flow dependency strength and the causal effect strength of the root cause anomaly source, a propagation attenuation factor is calculated to obtain the propagation causal effect strength of each successor node. Successor nodes whose propagation causal effect strength exceeds a preset propagation threshold are marked as propagation nodes and the root cause anomalies are merged to construct a propagation node sequence. The propagation causal effect intensity of each node in the propagation node sequence is sorted in descending order and the warning level division interval is determined. The propagation causal effect intensity is mapped to the warning level division interval to obtain the warning level identifier of each node. Based on the warning level identifier and the propagation causal effect intensity, a graded warning instruction is generated and sent to the device control layer corresponding to each node for execution.

[0012] A second aspect of the present invention provides a controller production equipment status early warning system, comprising: The data encoding unit is used to acquire multimodal time-series data of production equipment operation through edge sensing nodes, construct equipment topology structure according to the process flow dependency relationship between production equipment, and encode the multimodal time-series data based on the equipment topology structure to obtain equipment state representation vector. An anomaly detection unit is used to perform probability distribution modeling on the device state representation vector to generate a state prior distribution and a state posterior distribution, calculate the divergence value between the state prior distribution and the state posterior distribution as the anomaly confidence, and mark the device state representation vector whose anomaly confidence exceeds an adaptive threshold as an anomaly state representation vector. The root cause tracing unit is used to identify related nodes with causal dependencies by using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, and by using a causal graph inference algorithm to identify related nodes with causal dependencies along the device topology. It extracts the historical device state representation vectors of the related nodes, performs counterfactual reasoning on the historical device state representation vectors to generate comparative state representation vectors, and calculates the causal effect strength between the historical device state representation vectors and the comparative state representation vectors to identify the root cause anomaly source. The early warning control unit is used to construct an early warning propagation path along the device topology based on the root cause anomaly source, generate hierarchical early warning instructions according to the causal effect intensity of each node in the early warning propagation path, and send them to the device control layer for execution.

[0013] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] In this invention, multimodal time-series data collected by edge sensing nodes is encoded with equipment topology and then integrated with process flow dependencies, significantly improving the completeness and accuracy of abnormal state characterization. Probability distribution modeling simultaneously generates prior and posterior state distributions, quantifying anomaly confidence using divergence values. Adaptive threshold dynamic adjustment ensures a balance between anomaly labeling sensitivity and specificity, greatly reducing the risk of false alarms and missed alarms. Causal graph inference traces back along the topology from the anomaly timestamp, and counterfactual reasoning generates comparative state characterization vectors to directly remove interfering factors. Accurate causal effect strength calculation precisely locates the root cause of the anomaly, avoiding misjudgments caused by false causality in traditional correlation analysis. The early warning propagation path fully maps the process logic of anomaly spread, quantifying the impact of causal effects at each node. Hierarchical early warning commands enable differentiated responses, allowing the equipment control layer to perform targeted operations such as shutdown, speed reduction, or parameter correction, reducing unplanned downtime losses and delaying fault propagation. The entire process is driven by the inherent structure of the equipment, and its unsupervised nature avoids labeling dependencies, resulting in significantly lower deployment costs than supervised solutions. It is particularly suitable for building early warning systems in the early stages of production lines lacking historical fault data. The combination of tiered instructions and causal tracing not only prevents the spread of current anomalies but also provides root cause suggestions for equipment maintenance, forming a closed-loop decision-making chain from perception to control. This overall early warning mechanism drives the transformation of production equipment from passive maintenance to proactive defense, improving anomaly location efficiency by several orders of magnitude and providing reliable support for equipment health management in the industrial internet environment. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the controller-based production equipment status early warning method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the abnormal propagation node identification and hierarchical early warning process of the controller production equipment status early warning method according to an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the controller production equipment status early warning method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Multimodal time-series data of production equipment operation is acquired through edge sensing nodes, equipment topology is constructed according to the process flow dependency relationship between production equipment, and equipment state representation vector is obtained by encoding the multimodal time-series data based on the equipment topology. The device state representation vector is modeled with a probability distribution to generate a state prior distribution and a state posterior distribution. The divergence value between the state prior distribution and the state posterior distribution is calculated as the anomaly confidence. Device state representation vectors whose anomaly confidence exceeds the adaptive threshold are marked as anomaly state representation vectors. Using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, the causal graph inference algorithm is used to identify related nodes with causal dependencies along the device topology. The historical device state representation vectors of the related nodes are extracted, and counterfactual reasoning is performed on the historical device state representation vectors to generate comparative state representation vectors. The causal effect strength between the historical device state representation vectors and the comparative state representation vectors is calculated to identify the root cause of the anomaly. Based on the root cause anomaly source, an early warning propagation path is constructed along the device topology. According to the causal effect intensity of each node in the early warning propagation path, a graded early warning command is generated and sent to the device control layer for execution.

[0020] In one alternative implementation, Multimodal time-series data of production equipment operation is acquired through edge sensing nodes. An equipment topology is constructed based on the process dependencies between production equipment. The multimodal time-series data is then encoded based on the equipment topology to obtain an equipment state representation vector, including: Operational signals are collected by edge sensing nodes deployed at various production equipment. The operational signals are filtered and preprocessed, and the multimodal time series data is obtained synchronously according to the timestamp. The multimodal time series data is decomposed into time and frequency domain features. The multimodal time series data is statistically analyzed by sliding window to obtain trend features, which are then concatenated with the frequency domain features to obtain a multimodal feature vector. Extract the node identifiers and edge connections of each production device from the device topology, construct an adjacency matrix based on the edge connections, calculate the degree matrix of the adjacency matrix and normalize it to obtain the propagation matrix, and arrange the multimodal feature vectors according to the node identifiers to form a node feature matrix; The node feature matrix and the propagation matrix are multiplied to obtain the first-order neighborhood features. The first-order neighborhood features are nonlinearly activated and multiplied with the propagation matrix to obtain the second-order neighborhood features. The node feature matrix is ​​then weighted by residuals and linearly projected based on the second-order neighborhood features to obtain the device state characterization vector.

[0021] Edge sensing nodes deployed at various production equipment are responsible for real-time acquisition of equipment operation signals. These signals encompass various physical quantities such as vibration acceleration, current, temperature, pressure, and rotational speed, and are continuously output as raw time-series data by sensors at a fixed sampling frequency. Due to differences in sampling rates among different types of sensors, and the presence of electromagnetic interference and mechanical noise in industrial environments that introduce random noise and glitches into the signals, the raw operating signals require filtering preprocessing before entering the subsequent processing flow. The filtering preprocessing employs a combination of bandpass filtering and median filtering. Bandpass filtering removes high-frequency noise and low-frequency drift outside the frequency range, while median filtering eliminates pulse-like anomalies. After filtering, because the clock references of each edge sensing node may have slight deviations, the signals acquired by each node need to be aligned and synchronized using a unified global timestamp as a reference. This integrates the signals from each device at the same moment into multimodal time-series data, ensuring strict temporal correspondence between data channels during subsequent cross-device analysis.

[0022] Time-frequency decomposition is performed on multimodal time-series data. Short-time Fourier transform or wavelet transform is used to map the time-domain signal to the frequency domain, extracting frequency domain features such as energy distribution, dominant frequency amplitude, and spectral centroid of each frequency band. This reflects the changes in the operating status of equipment at different frequency scales; for example, bearing failures typically produce characteristic frequency components in specific frequency bands. Simultaneously, sliding window statistics are performed on the multimodal time-series data. A fixed-length sliding window slides along the time axis, calculating statistics such as mean, standard deviation, peak-to-peak value, skewness, and kurtosis within each window to form trend features reflecting the overall trend and fluctuation characteristics of the signal. The frequency domain features and trend features are concatenated along the feature dimension to obtain a multimodal feature vector that simultaneously contains frequency structure information and time trend information. Let the i-th... The multimodal feature vector of the device is ,in This represents the total dimension of the concatenated features. For device node indexing.

[0023] A topology is constructed based on the process dependencies between production equipment. Each piece of equipment is considered a node in the graph, and connections between equipment related to material flow, process sequence, or control signal transmission are considered edges. After extracting the node identifier set and edge connection set for each production equipment from the topology, an adjacency matrix is ​​constructed based on the edge connection relationships. ,in The total number of device nodes, if the number of nodes With nodes If there are connecting edges Otherwise To ensure that nodes consider their own state when aggregating neighborhood information, a self-loop is added to the adjacency matrix, resulting in... ,in It is an identity matrix.

[0024] calculate degree matrix ,in Represents a node The degree value (i.e., the sum of the number of edges connected to it, including self-loops). The adjacency matrix is ​​symmetrically normalized to obtain the propagation matrix. Normalization eliminates the impact of node degree differences on the feature aggregation amplitude, preventing high-degree nodes from imposing excessive weights on neighborhood features during aggregation. The multimodal feature vectors of each device node are arranged row-by-row according to node identifier order to form a node feature matrix. The first of the matrix The row corresponds to the first Multimodal feature vectors of the device .

[0025] In the graph convolutional encoding stage, the node feature matrix and propagation matrix Perform matrix multiplication to obtain the first-order neighborhood features. ,in This is the first layer learnable weight matrix. This represents the hidden layer feature dimension. First-order neighborhood features aggregate the feature information of each node and its directly neighboring device nodes, reflecting the impact of local process dependencies on the device state. Nonlinear activation is performed using the ReLU activation function. Introducing nonlinear expressive power to obtain activated features Then combine it with the propagation matrix Perform matrix multiplication to obtain the second-order neighborhood features. ,in This is the second-layer learnable weight matrix. This is the output feature dimension. Second-order neighborhood features further expand the receptive field based on first-order aggregation, enabling each device node to perceive the state information of neighboring devices within a distance of 2 hops, capturing a wider range of process dependencies.

[0026] To avoid oversmoothing issues caused by deep graph convolutions and to preserve the original feature information of the nodes, residual weighting and linear projection are performed on the node feature matrix based on second-order neighborhood features. Specifically, the node feature matrix... Through linear projection matrix Mapping to In the same dimensional space, the residual terms are obtained. Then, with learnable scalar weights The residual terms are weighted and added to the second-order neighborhood features to obtain the fused feature representation. Ultimately, for Perform linear projection using the projection matrix. Adjust the feature dimension to the target representation dimension. The device state representation vector matrix is ​​obtained. , of which OK That is, the first The device state representation vector of each piece of equipment. This vector simultaneously encodes the multimodal operating characteristics of the equipment itself and the state information of neighboring equipment transmitted through the process topology, providing a topology-aware feature foundation for subsequent probability distribution modeling and anomaly detection.

[0027] In practical engineering deployments, the aforementioned graph convolutional encoding process is executed periodically in sliding time windows. Within each time window, the device status representation vector of each device node is updated once, thus forming a continuous device status monitoring stream. After completing filtering and synchronization, the edge sensing nodes upload the data to the edge computing layer, where feature extraction and graph convolutional encoding are performed. This reduces data transmission bandwidth requirements and cloud computing latency, meeting the real-time requirements of industrial production.

[0028] In one alternative implementation, The device state representation vector is modeled using probability distribution to generate a prior state distribution and a posterior state distribution. The divergence value between the prior state distribution and the posterior state distribution is calculated as the anomaly confidence score. Device state representation vectors whose anomaly confidence scores exceed an adaptive threshold are marked as anomaly state representation vectors. The device state representation vector is nonlinearly mapped to obtain a latent space representation. Based on the latent space representation, a distribution mean and distribution variance are generated through dual-path encoding. A state posterior distribution is constructed based on the distribution mean and distribution variance. After performing a log-space transformation on the distribution variance, a noise vector is sampled from the standard normal distribution and a posterior sampling vector is constructed through reparameterization. The posterior sampling vector is reconstructed through multi-level decoding to obtain a reconstructed representation vector. The reconstruction loss between the device state representation vector and the reconstructed representation vector is calculated and a state prior distribution is constructed through maximum likelihood estimation. The divergence between the posterior and prior state distributions is calculated as the anomaly confidence level. Quantile statistics are performed on the anomaly confidence levels of each device node within a pre-stored historical sliding time window to obtain distribution feature parameters. A dynamic adjustment factor is calculated based on the distribution feature parameters and the current device load state. An adaptive threshold is obtained by adjusting a preset baseline threshold using the dynamic adjustment factor. The anomaly confidence level is compared with the adaptive threshold. Device state representation vectors whose anomaly confidence levels exceed the adaptive threshold are marked as anomaly state representation vectors.

[0029] After obtaining the equipment state representation vector, probability distribution modeling is performed. The equipment state representation vector is input into a nonlinear mapping module composed of a multi-layer fully connected network. Through layer-by-layer linear transformations and activation functions, the original equipment state representation vector is compressed and mapped to a low-dimensional latent space, resulting in a latent space representation. This nonlinear mapping process effectively removes redundant information from the equipment state representation vector, retaining the essential features closely related to the equipment's operating state, providing a more compact and discriminative input for subsequent probability modeling.

[0030] Based on the latent space representation, the distribution mean and distribution variance are generated separately through a dual-path encoding structure. Specifically, two independent fully connected branches take the latent space representation as input, and one branch outputs the distribution mean vector. The other branch outputs the distribution variance vector. The parameters of the two branches are independent, learning the desired location and uncertainty of the device state respectively. Based on the distribution mean... With distribution variance The posterior distribution of the state is modeled as a multivariate Gaussian distribution. ,in As latent variables, The input is the device state representation vector.

[0031] To make the sampling process of the posterior distribution differentiable to support end-to-end training, the variance of the distribution... Perform a logarithmic space transformation to obtain the logarithmic variance. This maps the variance's numerical range to the logarithmic domain, avoiding the problem of negative variance or numerical explosion during training. Subsequently, noise vectors are independently sampled from the standard normal distribution. ,in To define the latent spatial dimension, randomness is separated from the sampling operation through reparameterization techniques, and a posterior sampling vector is constructed. ,in , This indicates element-wise multiplication. This reparameterization method allows gradients to be successfully backpropagated to the mean and variance branches, ensuring the trainability of the overall framework.

[0032] For the posterior sampled vector The input is processed by a multi-layer decoding network, which uses layer-by-layer upsampling and nonlinear activation to reconstruct the representation in the latent space back to a reconstructed representation vector with the same dimension as the original device state representation vector. The number of layers in the decoding network is designed to be symmetrical with that of the encoding network to ensure sufficient expressive power during the reconstruction process. The original device state representation vector is then calculated. With reconstructed representation vector The reconstruction loss between the two is measured using the mean squared error method, i.e. Within the maximum likelihood estimation framework, the reconstruction loss corresponds to the negative log-likelihood of a Gaussian likelihood function with the reconstructed representation vector as the mean and a fixed variance as the parameter, thus modeling the state prior distribution as a standard normal distribution. This serves as a regularization reference for the posterior distribution. The training objective of the entire variational autoencoder framework is a weighted sum of the reconstruction loss and the distribution divergence loss. By minimizing this joint objective, the posterior distribution of the encoder output is made to converge towards the prior distribution while fitting the input data.

[0033] Calculate the posterior distribution of the state With state prior distribution The KL divergence between them serves as the anomaly confidence level for the current device state. The calculation form is For the aforementioned Gaussian assumption, the KL divergence has an analytical form: subscript For latent spatial dimension indexing, denoted as the total dimension of the potential space. When the equipment is operating normally, the posterior distribution and the prior distribution highly overlap, and the KL divergence value is small. When the equipment malfunctions, the mean shift or variance of the posterior distribution increases significantly, leading to a significant increase in the KL divergence, thus generating a higher anomaly confidence level.

[0034] Static thresholds are difficult to adapt to fluctuations in equipment load and changes in operating conditions in actual production scenarios. Therefore, an adaptive threshold mechanism is adopted to dynamically judge the anomaly confidence level. Quantile statistics are performed on the anomaly confidence level sequence of each device node within a pre-stored historical sliding time window, and high quantile values ​​(such as the 95th percentile) are extracted. With median As distribution characteristic parameters, these two parameters respectively reflect the extreme and typical levels of historical anomaly confidence, jointly characterizing the location and dispersion of the historical distribution. Simultaneously, the current equipment load status indicators are obtained. This indicator can be obtained by normalizing and fusing multi-dimensional signals such as power, speed, and capacity utilization rate collected by the equipment in real time, and its value range is [value range missing]. The higher the value, the higher the current load on the device.

[0035] Calculate the dynamic adjustment factor based on the distribution characteristic parameters and the current equipment load status. Its form is ,in and The preset regularization weight coefficients control the contribution of historical distribution dispersion and real-time load to threshold adjustment, respectively. When the historical anomaly confidence level of the device fluctuates significantly or the current load is high, the dynamic adjustment factor... When the value is large, the adaptive threshold should be adjusted upwards to reduce false alarms under high load conditions; conversely, when historical anomaly confidence is concentrated and the load is low, the threshold should be tightened to improve the detection sensitivity of subtle anomalies. This is achieved by dynamically adjusting the factor. For the preset benchmark threshold Adjustments are made to obtain an adaptive threshold. ,in It was determined by offline historical data statistics and served as the initial reference benchmark.

[0036] The anomaly confidence level of the current device node. With the corresponding adaptive threshold Compare. If If so, the device state representation vector corresponding to that device node is marked as an abnormal state representation vector, and the corresponding timestamp and device node identifier are recorded for subsequent cause-effect graph inference and root cause tracing; if If the condition is not met, the device is considered to be in normal operating condition, and no further warning process will be triggered. Through the synergistic effect of the above-mentioned probabilistic modeling and adaptive threshold mechanism, accurate and robust identification of abnormal device states can be achieved under different operating conditions and load conditions, effectively reducing the false alarm rate and the missed alarm rate, and laying a reliable foundation for subsequent causal inference and hierarchical early warning.

[0037] In one alternative implementation, Using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, the causal graph inference algorithm is used to identify associated nodes with causal dependencies along the device topology, and the historical device state representation vectors of the associated nodes are extracted, including: Node identifiers are extracted from the abnormal state representation vectors. The incoming edge connections of the node identifiers are extracted from the device topology and a candidate predecessor node set is determined. The historical device state representation vectors of each node in the candidate predecessor node set are extracted and the temporal Granger causality test statistic is calculated. Nodes with causal precedence relationships are selected based on the Granger causality test statistic to obtain a first-order causal predecessor node set. For each node in the first-order causal predecessor node set, a predecessor node extraction operation is recursively performed. In each recursion, the incoming edge connection relationship of the current node is extracted to obtain a recursive candidate predecessor node set and the corresponding historical device state representation vector is determined. Based on the historical device state representation vector, the time-series Granger causality test statistic is calculated and indirect causal relationships are eliminated through conditional independence test to obtain multiple direct causal predecessor nodes and construct a multi-order causal predecessor node set. When the recursion depth reaches a preset number of layers, the recursion is terminated and the first-order causal predecessor node set and the multi-order causal predecessor node set are merged to obtain an associated node set. Using the Granger causality test statistic as edge weights, a directed weighted causal graph is constructed between the set of associated nodes and the abnormal nodes corresponding to the abnormal state representation vector. Based on the directed weighted causal graph, the historical device state representation vector of each node in the set of associated nodes is extracted.

[0038] After labeling the abnormal state representation vectors, it is necessary to perform source analysis to determine the root cause of the current equipment malfunction. Using the timestamp corresponding to the abnormal state representation vector as the starting point, node identifiers are extracted from the vectors. Each node identifier uniquely corresponds to a specific equipment node in the equipment topology and is designated as the abnormal node. Subsequently, all incoming edge connections of this node identifier are retrieved in the equipment topology. Incoming edges represent adjacent nodes with upstream output relationships in the process flow. All nodes directly connected to the abnormal node through incoming edges are included in the candidate predecessor node set. Nodes in the candidate predecessor node set are located directly upstream of the abnormal node in the process flow, possessing the physical basis for propagating state disturbances downstream.

[0039] For each node in the candidate predecessor node set, a sequence of historical device state representation vectors within a historical window prior to the traced timestamp is extracted, and the time-series Granger causality test statistic is calculated. The core idea of ​​the Granger causality test is: if, given the known historical state of an anomalous node, adding the historical state sequence of a candidate predecessor node significantly reduces the prediction error of the anomalous node's current state, then the candidate predecessor node is considered to have a Granger causal lead relationship with the anomalous node. Specifically, for each candidate predecessor node, both a restricted autoregressive model and an unrestricted autoregressive model are constructed, and the F-statistic corresponding to the ratio of the sum of squared residuals of the two models is calculated. ,in Indicates candidate predecessor nodes For abnormal nodes The Granger causality test statistic. A significance level is set. ,when The corresponding p-value is lower than At that time, determine the candidate predecessor node. with abnormal nodes A significant causal precedence relationship exists between the nodes, and these nodes are included in the set of first-order causal precursor nodes. After screening, the set of first-order causal precursor nodes retains upstream nodes that have a direct statistical causal relationship with the anomalous nodes, while excluding nodes that are only adjacent in process location but do not actually have a state transmission relationship.

[0040] After determining the set of first-order causal precursor nodes, a precursor node extraction operation is recursively performed on each node to expand the causal tracing range upstream layer by layer. In each recursion, all incoming edge connections of the current node in the device topology are taken to obtain a recursive candidate precursor node set, and the historical device state representation vector sequence within the corresponding time window of each node in the recursive candidate precursor node set is extracted. Based on these historical device state representation vector sequences, the time-series Granger causality test statistic is calculated again to screen out nodes with Granger causal precedence relationships with the current node. However, the causal relationships identified solely by the Granger causality test may contain indirect causal paths, meaning that the influence of a node on the current node is actually transmitted through intermediate nodes rather than acting directly. Therefore, a conditional independence test is introduced to further purify the screening results: for each pair of nodes that pass the Granger test, given the historical state sequence of their common neighbor nodes as a condition variable, the partial correlation between the two nodes is tested to see if it is still significant. If the conditional independence test results indicate that the two are no longer significantly correlated after controlling for intermediate nodes, then the causal relationship is determined to be indirect and is removed, retaining only the direct causal precursor nodes. The direct causal precursor nodes obtained in each round of recursion are accumulated and added to the multi-level causal precursor node set. The recursion depth is determined by a preset number of levels. Control, when the recursion depth reaches The recursion terminates at this point, merging the set of first-order causal predecessor nodes with the set of multi-order causal predecessor nodes to obtain the complete set of associated nodes. Preset number of layers. The settings need to take into account the actual transmission depth of the process flow and the computational overhead. They are usually configured according to the number of process levels in the production line to avoid introducing remote nodes that are not related to the current anomaly due to excessive recursion.

[0041] After the set of associated nodes is constructed, the Granger causality test statistic is used. As weights for directed edges, a directed weighted causal graph is constructed between each node in the set of associated nodes and the anomalous node. Each directed edge in the directed weighted causal graph points from the upstream causal predecessor node to the downstream affected node. The edge weight reflects the statistical strength of the corresponding causal relationship; a larger weight indicates a more significant influence of the upstream node on the state of the downstream node. For multi-order causal paths, the directed weighted causal graph contains directed paths spanning multiple process levels. The weights of each edge on the path collectively describe the cumulative effect of state disturbances propagating along the process flow. The construction of the directed weighted causal graph organizes the discrete set of associated nodes into a structured causal network with clear directionality and strength information, providing a topological foundation for subsequent counterfactual reasoning and root cause localization.

[0042] Based on the constructed directed weighted causal graph, historical device state representation vectors are extracted from each node in the set of associated nodes. The extraction range is based on the timestamps of anomalous nodes, tracing back a time window. The length of the time window is consistent with the historical window used in the Granger causality test to ensure that the extracted historical device state representation vectors are consistent with the data on which the causality test depends in terms of time range. The extracted historical device state representation vectors are organized by node index and time step to form a historical state matrix of associated nodes, which is used in subsequent counterfactual reasoning steps.

[0043] In one alternative implementation, Performing counterfactual reasoning on the historical equipment state representation vector to generate a contrasting state representation vector, and calculating the causal effect strength between the historical equipment state representation vector and the contrasting state representation vector to identify root cause anomalies includes: The causal paths between each associated node and the abnormal node are extracted from the directed weighted causal graph, and the causal strength of the path is calculated based on the edge weight. The associated nodes whose causal strength of the path exceeds a preset threshold are selected as the intervention candidate node set. The device state representation vector corresponding to the intervention candidate node set is extracted from the historical device state representation vector as the original state representation vector. The original state representation vector is variationally encoded and mapped to the latent space to obtain latent distribution parameters. Abnormal latent dimensions in the latent distribution parameters are identified, truncated, and resampled to obtain a corrected latent vector. The corrected latent vector is decoded, mapped, and reconstructed to the original representation space to obtain an intervention state representation vector. The intervention state representation vector is used as the initial input and propagated along the causal path using graph convolution. At each propagation node, the input representation vector is weighted and aggregated according to the edge weights, and the node state representation is updated. The propagation is then performed to the abnormal node position to obtain a comparison state representation vector. The representation space distance between the abnormal state representation vector and the comparison state representation vector is calculated as the counterfactual bias. The counterfactual bias and the path causal intensity of the corresponding node are weighted and normalized to obtain the causal effect intensity. The node with the largest causal effect intensity is selected as the root cause anomaly source.

[0044] After extracting the historical state representations of the associated nodes, it is necessary to further quantify the actual causal contribution of each candidate node to the abnormal event through counterfactual reasoning mechanisms, thereby accurately locating the root cause of the anomaly. This process unfolds based on a directed weighted causal graph. In the directed weighted causal graph, each directed edge represents the direction of the causal influence of one device node on another device node, and the weight value attached to the edge reflects the historical statistical strength of the causal influence, which is usually obtained by normalizing the F-statistic or partial correlation coefficient of the Granger causality test.

[0045] Extract all directed paths from each associated node to an anomalous node from the directed weighted causal graph. For each path, the path causality strength is obtained by multiplying the weights of all edges along the path. If there are multiple paths between an associated node and an anomalous node, the maximum value of all path causality strengths is taken as the comprehensive path causality strength of that node. Let the path... The above contains A directed edge, the first The weight of the edge is Then the path Path causality strength for After taking the maximum value for all paths, compare it with a preset threshold. Comparison, only retaining paths with causal strength exceeding The associated nodes constitute the intervention candidate node set. Noise nodes with extremely low weight and negligible actual contribution in the causal propagation chain are excluded, thus focusing subsequent counterfactual reasoning calculations on high-confidence candidate sources. From the extracted historical device state representation vectors, the corresponding device state representation vectors are extracted according to the member index of the intervention candidate node set, and used as the original state representation vector input for subsequent variational coding.

[0046] The original state representation vector is input into the encoder part of a pre-trained variational autoencoder, which maps it to the latent space through a multi-layer feedforward network in the encoder, and outputs the mean vector of the latent distribution. With log-variance vector These two together constitute the latent distribution parameters. Each dimension in the latent space corresponds to an abstract feature component of the device state. Anomaly detection is performed on the distribution parameters of each dimension: if the mean of a certain dimension... (in If the value of a latent dimension (as an index for a potential dimension) exceeds the upper or lower bounds of the confidence interval obtained statistically from normal historical samples, then the dimension is identified as an anomalous latent dimension. For identified anomalous latent dimensions, a truncation and resampling operation is performed—the mean of the dimension is truncated to the boundary value of the normal confidence interval. The variance remains unchanged, and samples are resampled within the truncated distribution range to obtain the corrected sampled values ​​for that dimension. The sampled values ​​for all dimensions (normal dimensions sampled according to the original parameters, and abnormal dimensions sampled according to the truncated parameters) are concatenated to form the corrected latent vector. This truncation and resampling process is semantically equivalent to a counterfactual intervention operation that "assumes that the abnormal latent features of the node are corrected to a normal boundary state," and is the core step of counterfactual reasoning.

[0047] Will correct the latent vector The input is fed into the decoder of the variational autoencoder. The decoder reconstructs the latent vector back to the original representation space dimension through a multi-layer feedforward network, obtaining the intervention state representation vector. The intervention state representation vector represents the device state representation of the candidate node under the assumption that its state is "restored" to the normal boundary condition. As initial input, graph convolution propagation is performed along the causal path from the candidate node to the anomaly node in the directed weighted causal graph. During propagation, at each intermediate node, the incoming representation vector is weighted and aggregated according to the weight of the incident edge, and then proportionally merged with the historical state representation vector of that intermediate node to update the propagation state representation at that node. Let the incoming representation vector received by the current propagation node from the predecessor node be... The weight of this edge is The historical state of this intermediate node is represented as follows: The updated node propagation state representation for ,in The learnable weight matrix for the graph convolutional propagation layer. The ReLU nonlinear activation function definition is adopted as described above. Propagation proceeds node by node according to this rule until the position of the anomaly node is reached, obtaining the contrastive state representation vector corresponding to that candidate node. The contrast state representation vector semantically represents the state representation that should theoretically be presented at the abnormal node if the abnormal potential features of the candidate node are intervened and corrected.

[0048] Calculate the abnormal state representation vector Compared with the state representation vector The Euclidean distance between them serves as a counterfactual bias. ,Right now The larger the counterfactual bias, the more significant the impact of the candidate node's anomalous state on the anomalous node, indicating a stronger causal contribution from the candidate node. The counterfactual bias... Comprehensive path causality strength corresponding to the candidate node The final causal effect strength is obtained by normalization through weighted fusion. The specific calculation method is to perform a calculation on all candidate intervention nodes. and Perform min-max normalization separately, then use the weighted coefficients. and (satisfy Perform linear fusion: ,in and These are the normalized counterfactual bias and the path causality strength, respectively. This fusion design takes into account two perspectives: the path causality strength reflects the prior propagation weights at the causal graph structure level, while the counterfactual bias reflects the intervention response magnitude at the actual data distribution level. The fusion of the two can effectively avoid the risk of misjudgment caused by a single indicator.

[0049] Iterate through all intervention candidate nodes and calculate the causal effect strength of each node. The node with the strongest causal effect is selected as the root cause anomaly source. This node is identified as the origin of the anomaly in the device topology, and its abnormal state propagates through the causal path, ultimately causing the state deviation of the target anomaly node. The identified root cause anomaly source information will serve as the input for subsequent early warning propagation path construction and hierarchical early warning command generation, ensuring that the early warning response accurately points to the actual source of the fault rather than intermediate propagation nodes.

[0050] In one alternative implementation, The process involves identifying anomalous latent dimensions in the latent distribution parameters, truncating and resampling them to obtain a corrected latent vector, and then decoding and mapping the corrected latent vector back to the original representation space to obtain the intervention state representation vector, including: Extract the distribution mean and distribution variance of each potential dimension from the potential distribution parameters, calculate the deviation of the distribution variance of the potential dimension from the distribution variance of the corresponding dimension in the pre-trained normal state potential space, mark the potential dimensions with the deviation exceeding the preset deviation threshold as abnormal potential dimensions, and extract the corresponding distribution mean and distribution variance to construct an abnormal dimension distribution set. Extract the mean and variance of the normal distribution for each position of the abnormal dimension in the abnormal dimension distribution set from the pre-trained normal state latent space. Construct a normal dimension distribution based on the mean and variance of the normal distribution and perform random sampling to obtain normal dimension sample values. Replace the distribution parameters of the abnormal latent dimension in the latent distribution parameters with the normal dimension sample values ​​to obtain the corrected distribution parameters. Based on the modified distribution parameters, a modified latent vector is constructed and mapped to a high-dimensional feature space through multi-layer nonlinear decoding to obtain intermediate reconstructed features. The intermediate reconstructed features are then dimensionally expanded and mapped to the representation dimension corresponding to the original state representation vector through a linear projection layer to obtain the intervention state representation vector.

[0051] After encoding the latent distribution parameters of the historical states of associated nodes, it is necessary to perform dimension-by-dimensional anomaly identification on the latent space of the encoded output. This allows the distribution parameters of the anomalous dimensions to be replaced with the corresponding parameters under normal conditions, thereby constructing an intervention representation that reflects "if the device were in a normal state." Instead of replacing the entire latent vector, this process precisely locates the potentially shifted dimensions, truncates and resamples only these dimensions, and retains the original encoded information of the remaining dimensions, ensuring the minimization and interpretability of the intervention operation.

[0052] The latent distribution parameters output by the encoder include the distribution mean for each latent dimension. With distribution variance Extract the variance values ​​of each dimension one by one. (in (For latent dimension indexing). During the pre-training phase, the variational autoencoder is trained using a large amount of historical device data under normal operating conditions. After training convergence, the statistical parameters of each dimension in the normal state latent space are saved, including the mean of the normal distribution. variance compared to normal distribution Using the aforementioned pre-training parameters as a reference, the deviation between the variance of each latent dimension of the current candidate node's encoded output and the variance of the corresponding dimension in the normal state is calculated. The calculation method is as follows That is, the relative deviation ratio with the normal variance as the denominator. When Exceeding the preset deviation threshold At that time, the first Each latent dimension is labeled as an abnormal latent dimension, and its corresponding distribution mean is calculated. With distribution variance Collected into the set of abnormal dimensional distributions .

[0053] Anomaly Dimensional Distribution Set After construction, extract from the pre-trained normal state latent space The mean of the normal distribution at each position corresponding to the anomaly dimension. variance compared to normal distribution Based on these two parameters, construct the normal distribution for this dimension. Random sampling is performed on this distribution to obtain normal dimension sample values. The sampling process employs a reparameterization technique, that is, letting... ,in This represents the noise components independently sampled from the standard normal distribution. For those not belonging to... The normal latent dimension preserves the distribution mean of the original encoded output. This serves as the sample value for that dimension. The sample values ​​from all dimensions are concatenated to obtain the corrected distribution parameter vector. In this case, the abnormal dimension has been replaced with the sampled value under normal conditions, while the normal dimension retains the original encoded information.

[0054] Based on the modified distribution parameter vector A modified latent vector is constructed and input into a multi-layer nonlinear decoding network for mapping. The decoding network consists of multiple fully connected layers and nonlinear activation functions stacked alternately. After performing an affine transformation, each layer undergoes nonlinear activation such as ReLU or Tanh, progressively mapping the modified representation in the low-dimensional latent space to the high-dimensional feature space to obtain intermediate reconstructed features. The dimensions of intermediate reconstructed features are typically set to match the output dimensions of the encoder's intermediate layers to ensure symmetry between the decoding and encoding paths, thereby enhancing the information recovery capability of the reconstruction process. The number of layers and hidden layer widths of the decoding network are fixed during model pre-training; during inference, pre-trained weights are directly loaded for forward computation, eliminating the need for additional training.

[0055] Obtaining intermediate reconstruction features Next, dimensionality expansion is performed. This dimensionality expansion operation is implemented through a linear projection layer, whose weight matrix... Reconstructing intermediate features from dimensions Target dimension mapped to the original state representation vector The calculation process is as follows ,in This is the bias vector. The obtained... This is the intervention state representation vector, whose dimension is the same as the original device state representation vector. Completely consistent, and can be directly used in subsequent counterfactual deviation calculations.

[0056] Intervention state representation vector The physical meaning of is: it represents the state that a node should exhibit in the original representation space after forcibly replacing the abnormal potential dimensions of a candidate associated node with normal state parameters. This is achieved by... The warning propagates along the warning propagation path to the target anomaly node, obtaining the comparison state representation vector. Then, compared with the actual observed abnormal state representation vector By comparing the results, the actual contribution of anomalies in candidate nodes to the abnormal state of the target node can be quantified. The advantage of this truncated resampling mechanism is that it avoids the problem of excessive intervention caused by global replacement of the entire potential vector, and can achieve accurate counterfactual simulation of specific anomaly sources with minimal information disturbance, thereby improving the accuracy and reliability of root cause identification.

[0057] Preset deviation threshold The setting of this parameter directly affects the sensitivity of anomaly dimension identification. If... Setting it too small may misjudge normal statistical fluctuations as abnormal dimensions, leading to an excessively large intervention range; if Setting the value too high may miss true anomalies, resulting in insufficient intervention. In engineering practice, The root cause can be determined by performing a grid search on the validation set with the root cause identification accuracy as the optimization objective, or by using the historical distribution of variance of each dimension under normal conditions and the confidence interval boundary of the statistical significance test as the basis for automatic setting, thereby reducing the reliance on manual parameter tuning and improving the adaptability of the method.

[0058] Figure 2 This is a flowchart illustrating the abnormal propagation node identification and hierarchical early warning process of the controller production equipment status early warning method according to an embodiment of the present invention.

[0059] In one alternative implementation, Based on the root cause anomaly source, an early warning propagation path is constructed along the device topology. A tiered early warning command is generated according to the causal effect intensity of each node in the early warning propagation path and sent to the device control layer for execution, including: The outgoing edge connections of the root cause anomaly source are extracted from the device topology to obtain a set of successor nodes. The process flow dependency strength between each successor node and the root cause anomaly source is extracted. Based on the process flow dependency strength and the causal effect strength of the root cause anomaly source, a propagation attenuation factor is calculated to obtain the propagation causal effect strength of each successor node. Successor nodes whose propagation causal effect strength exceeds a preset propagation threshold are marked as propagation nodes and the root cause anomalies are merged to construct a propagation node sequence. The propagation causal effect intensity of each node in the propagation node sequence is sorted in descending order and the warning level division interval is determined. The propagation causal effect intensity is mapped to the warning level division interval to obtain the warning level identifier of each node. Based on the warning level identifier and the propagation causal effect intensity, a graded warning instruction is generated and sent to the device control layer corresponding to each node for execution.

[0060] After identifying the root cause anomaly, the outgoing edges of the equipment are traversed downstream along the equipment topology to extract all successor nodes directly connected to the root cause anomaly, forming a set of successor nodes. Each outgoing edge corresponds to a process dependency relationship, and the strength of this dependency relationship reflects the degree of process influence of upstream equipment on downstream equipment. The strength of the process dependency relationship can be quantified based on the frequency of coordinated changes in the states of upstream and downstream equipment, material transfer volume, or energy coupling degree in historical process data, with the value range normalized to [value range missing]. The larger the value in the range, the tighter the process coupling between upstream and downstream, and the higher the possibility of abnormal propagation.

[0061] For each successor node in the successor node set, the propagation attenuation factor is calculated by combining the process dependency strength with the root cause anomaly source and the causal effect strength of the root cause anomaly source itself. Let the causal effect strength of the root cause anomaly source be... successor node The process dependency strength between the source of the root cause anomaly and the source of the anomaly is The attenuation coefficient is Then the successor node The strength of the propagation causal effect Calculate using the following formula: ,in This is a preset propagation attenuation coefficient used to simulate the physical law of abnormal signals gradually attenuating in the process flow. Attenuation coefficient Offline calibration can be performed based on equipment type, process medium characteristics, or historical propagation statistics. Different configurations can be set for different process stages. The values ​​are selected to adapt to diverse production process scenarios.

[0062] The strength of the propagation causal effect of each successor node With preset propagation threshold Compare, only those that meet the requirements The successor node is marked as a propagation node. A preset propagation threshold is set. The threshold setting needs to comprehensively consider the response capability of the device control layer and the false alarm rate control requirements. Too low a threshold will cause a large number of low-risk nodes to trigger warnings, resulting in wasted resources; too high a threshold may miss real propagation risks. In actual deployment, Adaptive calibration can be performed by combining historical early warning data with receiver operating characteristic (ROC) curve analysis. All propagation nodes meeting the criteria are merged with the root cause anomaly source and arranged in topological order to construct a propagation node sequence. This sequence fully describes the impact range of this anomaly event in the production equipment network, providing a complete list of nodes for the generation of subsequent tiered early warning commands.

[0063] After the propagation node sequence is constructed, the propagation causal effect intensity values ​​of all nodes in the sequence are sorted in descending order. The warning level division intervals are determined based on the distribution characteristics of the intensity values. Warning levels are typically divided into three levels: high-risk (Level 1 warning), medium-risk (Level 2 warning), and low-risk (Level 3 warning). The boundary values ​​for dividing the intervals are then defined. and These correspond to the dividing points between high and medium risk, and between medium and low risk, respectively. The boundary values ​​can be determined by using a geometrical division or by using quantiles based on the historical propagation intensity distribution, for example, using the 75th percentile of the propagation causal effect intensity distribution as... Using the 40th percentile as This ensures that the number of early warning nodes at all levels maintains a reasonable statistical proportion, avoiding an excessive concentration of nodes at a certain level.

[0064] The strength of the propagation causal effect of each node in the propagation node sequence. Mapping to the warning level division interval yields the corresponding warning level identifier. When When, the node is marked as a Level 1 warning; when When, the node is marked as a level 2 warning; when At this point, the node is marked as a Level 3 warning. Root cause anomalies, due to their typically highest causal effect strength, are usually directly marked as Level 1 warnings. For nodes at the end of the propagation node sequence, having undergone multiple decays, their propagation causal effect strength often falls into the Level 3 warning range, and the corresponding control layer response measures are relatively mild, mainly consisting of encrypted monitoring and preventative checks.

[0065] Based on the warning level identifier and the strength of the propagation causal effect, a hierarchical warning instruction is generated for each node in the propagation node sequence. The first-level warning instruction includes a mandatory control command for immediate shutdown or reduced operation, fault isolation operation instructions, and maintenance resource scheduling requests. It also includes the propagation causal effect strength value of the current node and the root cause anomaly source identifier, which facilitates the equipment control layer operators to quickly locate the root cause of the problem. The second-level warning instruction includes a reduced operation suggestion, a key parameter encrypted collection instruction, and a backup equipment switching preparation command, requiring the equipment control layer to complete the status confirmation and feedback within a specified time window. The third-level warning instruction focuses on monitoring enhancement, instructing the equipment control layer to increase the sampling frequency, record detailed operation logs, and set local alarm flags for manual inspection reference.

[0066] After the tiered early warning commands are generated, they are distributed to the corresponding device control layers of each node via the industrial communication network. The distribution process employs a priority queue mechanism, with level 1 early warning commands having the highest transmission priority, ensuring they reach the target control layer first even under network congestion. Upon receiving the early warning command, the device control layer executes the corresponding control action based on the early warning level identifier and the strength of the propagation causal effect in the command, and reports the execution result to the upper-level early warning management platform, forming a closed-loop early warning response mechanism. For scenarios where multiple root cause anomalies are triggered concurrently, when merging the tiered early warning commands generated by each propagation path, duplicate commands from the same node must be deduplicated and their highest level selected to ensure that the final command received by each device control layer is the most stringent early warning level among all propagation paths, avoiding control layer response chaos due to command conflicts.

[0067] The entire process of constructing the early warning propagation path and generating instructions is calculated on the edge sensing node side. The sparsity of the device topology ensures computational efficiency and can complete the entire process from root cause identification to instruction issuance within the millisecond to second time range, meeting the timeliness requirements of real-time early warning for production equipment.

[0068] A second aspect of the present invention provides a controller production equipment status early warning system, comprising: The data encoding unit is used to acquire multimodal time-series data of production equipment operation through edge sensing nodes, construct equipment topology structure according to the process flow dependency relationship between production equipment, and encode the multimodal time-series data based on the equipment topology structure to obtain equipment state representation vector. An anomaly detection unit is used to perform probability distribution modeling on the device state representation vector to generate a state prior distribution and a state posterior distribution, calculate the divergence value between the state prior distribution and the state posterior distribution as the anomaly confidence, and mark the device state representation vector whose anomaly confidence exceeds an adaptive threshold as an anomaly state representation vector. The root cause tracing unit is used to identify related nodes with causal dependencies by using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, and by using a causal graph inference algorithm to identify related nodes with causal dependencies along the device topology. It extracts the historical device state representation vectors of the related nodes, performs counterfactual reasoning on the historical device state representation vectors to generate comparative state representation vectors, and calculates the causal effect strength between the historical device state representation vectors and the comparative state representation vectors to identify the root cause anomaly source. The early warning control unit is used to construct an early warning propagation path along the device topology based on the root cause anomaly source, generate hierarchical early warning instructions according to the causal effect intensity of each node in the early warning propagation path, and send them to the device control layer for execution.

[0069] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0070] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0071] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A controller-based early warning method for the status of production equipment, characterized in that, include: Multimodal time-series data of production equipment operation is acquired through edge sensing nodes, equipment topology is constructed according to the process flow dependency relationship between production equipment, and equipment state representation vector is obtained by encoding the multimodal time-series data based on the equipment topology. The device state representation vector is modeled with a probability distribution to generate a state prior distribution and a state posterior distribution. The divergence value between the state prior distribution and the state posterior distribution is calculated as the anomaly confidence. Device state representation vectors whose anomaly confidence exceeds the adaptive threshold are marked as anomaly state representation vectors. Using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, the causal graph inference algorithm is used to identify related nodes with causal dependencies along the device topology. The historical device state representation vectors of the related nodes are extracted, and counterfactual reasoning is performed on the historical device state representation vectors to generate comparative state representation vectors. The causal effect strength between the historical device state representation vectors and the comparative state representation vectors is calculated to identify the root cause of the anomaly. Based on the root cause anomaly source, an early warning propagation path is constructed along the device topology. According to the causal effect intensity of each node in the early warning propagation path, a graded early warning command is generated and sent to the device control layer for execution.

2. The method according to claim 1, characterized in that, Multimodal time-series data of production equipment operation is acquired through edge sensing nodes. An equipment topology is constructed based on the process dependencies between production equipment. The multimodal time-series data is then encoded based on the equipment topology to obtain an equipment state representation vector, including: Operational signals are collected by edge sensing nodes deployed at various production equipment. The operational signals are filtered and preprocessed, and the multimodal time series data is obtained synchronously according to the timestamp. The multimodal time series data is decomposed into time and frequency domain features. The multimodal time series data is statistically analyzed by sliding window to obtain trend features, which are then concatenated with the frequency domain features to obtain a multimodal feature vector. Extract the node identifiers and edge connections of each production device from the device topology, construct an adjacency matrix based on the edge connections, calculate the degree matrix of the adjacency matrix and normalize it to obtain the propagation matrix, and arrange the multimodal feature vectors according to the node identifiers to form a node feature matrix; The node feature matrix and the propagation matrix are multiplied to obtain the first-order neighborhood features. The first-order neighborhood features are nonlinearly activated and multiplied with the propagation matrix to obtain the second-order neighborhood features. The node feature matrix is ​​then weighted by residuals and linearly projected based on the second-order neighborhood features to obtain the device state characterization vector.

3. The method according to claim 1, characterized in that, The device state representation vector is modeled using probability distribution to generate a prior state distribution and a posterior state distribution. The divergence value between the prior state distribution and the posterior state distribution is calculated as the anomaly confidence score. Device state representation vectors whose anomaly confidence scores exceed an adaptive threshold are marked as anomaly state representation vectors. The device state representation vector is nonlinearly mapped to obtain a latent space representation. Based on the latent space representation, a distribution mean and distribution variance are generated through dual-path encoding. A state posterior distribution is constructed based on the distribution mean and distribution variance. After performing a log-space transformation on the distribution variance, a noise vector is sampled from the standard normal distribution and a posterior sampling vector is constructed through reparameterization. The posterior sampling vector is reconstructed through multi-level decoding to obtain a reconstructed representation vector. The reconstruction loss between the device state representation vector and the reconstructed representation vector is calculated and a state prior distribution is constructed through maximum likelihood estimation. The divergence between the posterior and prior state distributions is calculated as the anomaly confidence level. Quantile statistics are performed on the anomaly confidence levels of each device node within a pre-stored historical sliding time window to obtain distribution feature parameters. A dynamic adjustment factor is calculated based on the distribution feature parameters and the current device load state. An adaptive threshold is obtained by adjusting a preset baseline threshold using the dynamic adjustment factor. The anomaly confidence level is compared with the adaptive threshold. Device state representation vectors whose anomaly confidence levels exceed the adaptive threshold are marked as anomaly state representation vectors.

4. The method according to claim 1, characterized in that, Using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, the causal graph inference algorithm is used to identify associated nodes with causal dependencies along the device topology, and the historical device state representation vectors of the associated nodes are extracted, including: Node identifiers are extracted from the abnormal state representation vectors. The incoming edge connections of the node identifiers are extracted from the device topology and a candidate predecessor node set is determined. The historical device state representation vectors of each node in the candidate predecessor node set are extracted and the temporal Granger causality test statistic is calculated. Nodes with causal precedence relationships are selected based on the Granger causality test statistic to obtain a first-order causal predecessor node set. For each node in the first-order causal predecessor node set, a predecessor node extraction operation is recursively performed. In each recursion, the incoming edge connection relationship of the current node is extracted to obtain a recursive candidate predecessor node set and the corresponding historical device state representation vector is determined. Based on the historical device state representation vector, the time-series Granger causality test statistic is calculated and indirect causal relationships are eliminated through conditional independence test to obtain multiple direct causal predecessor nodes and construct a multi-order causal predecessor node set. When the recursion depth reaches a preset number of layers, the recursion is terminated and the first-order causal predecessor node set and the multi-order causal predecessor node set are merged to obtain an associated node set. Using the Granger causality test statistic as edge weights, a directed weighted causal graph is constructed between the set of associated nodes and the abnormal nodes corresponding to the abnormal state representation vector. Based on the directed weighted causal graph, the historical device state representation vector of each node in the set of associated nodes is extracted.

5. The method according to claim 1, characterized in that, Performing counterfactual reasoning on the historical equipment state representation vector to generate a contrasting state representation vector, and calculating the causal effect strength between the historical equipment state representation vector and the contrasting state representation vector to identify root cause anomalies includes: The causal paths between each associated node and the abnormal node are extracted from the directed weighted causal graph, and the causal strength of the path is calculated based on the edge weight. The associated nodes whose causal strength of the path exceeds a preset threshold are selected as the intervention candidate node set. The device state representation vector corresponding to the intervention candidate node set is extracted from the historical device state representation vector as the original state representation vector. The original state representation vector is variationally encoded and mapped to the latent space to obtain latent distribution parameters. Abnormal latent dimensions in the latent distribution parameters are identified, truncated, and resampled to obtain a corrected latent vector. The corrected latent vector is decoded, mapped, and reconstructed to the original representation space to obtain an intervention state representation vector. The intervention state representation vector is used as the initial input and propagated along the causal path using graph convolution. At each propagation node, the input representation vector is weighted and aggregated according to the edge weights, and the node state representation is updated. The propagation is then performed to the abnormal node position to obtain a comparison state representation vector. The representation space distance between the abnormal state representation vector and the comparison state representation vector is calculated as the counterfactual bias. The counterfactual bias and the path causal intensity of the corresponding node are weighted and normalized to obtain the causal effect intensity. The node with the largest causal effect intensity is selected as the root cause anomaly source.

6. The method according to claim 5, characterized in that, The process involves identifying anomalous latent dimensions in the latent distribution parameters, truncating and resampling them to obtain a corrected latent vector, and then decoding and mapping the corrected latent vector back to the original representation space to obtain the intervention state representation vector, including: Extract the distribution mean and distribution variance of each potential dimension from the potential distribution parameters, calculate the deviation of the distribution variance of the potential dimension from the distribution variance of the corresponding dimension in the pre-trained normal state potential space, mark the potential dimensions with the deviation exceeding the preset deviation threshold as abnormal potential dimensions, and extract the corresponding distribution mean and distribution variance to construct an abnormal dimension distribution set. Extract the mean and variance of the normal distribution for each position of the abnormal dimension in the abnormal dimension distribution set from the pre-trained normal state latent space. Construct a normal dimension distribution based on the mean and variance of the normal distribution and perform random sampling to obtain normal dimension sample values. Replace the distribution parameters of the abnormal latent dimension in the latent distribution parameters with the normal dimension sample values ​​to obtain the corrected distribution parameters. Based on the modified distribution parameters, a modified latent vector is constructed and mapped to a high-dimensional feature space through multi-layer nonlinear decoding to obtain intermediate reconstructed features. The intermediate reconstructed features are then dimensionally expanded and mapped to the representation dimension corresponding to the original state representation vector through a linear projection layer to obtain the intervention state representation vector.

7. The method according to claim 1, characterized in that, Based on the root cause anomaly source, an early warning propagation path is constructed along the device topology. A tiered early warning command is generated according to the causal effect intensity of each node in the early warning propagation path and sent to the device control layer for execution, including: The outgoing edge connections of the root cause anomaly source are extracted from the device topology to obtain a set of successor nodes. The process flow dependency strength between each successor node and the root cause anomaly source is extracted. Based on the process flow dependency strength and the causal effect strength of the root cause anomaly source, a propagation attenuation factor is calculated to obtain the propagation causal effect strength of each successor node. Successor nodes whose propagation causal effect strength exceeds a preset propagation threshold are marked as propagation nodes and the root cause anomalies are merged to construct a propagation node sequence. The propagation causal effect intensity of each node in the propagation node sequence is sorted in descending order and the warning level division interval is determined. The propagation causal effect intensity is mapped to the warning level division interval to obtain the warning level identifier of each node. Based on the warning level identifier and the propagation causal effect intensity, a graded warning instruction is generated and sent to the device control layer corresponding to each node for execution.

8. A controller-based production equipment status early warning system, used to implement the method described in any one of claims 1-7, characterized in that, include: The data encoding unit is used to acquire multimodal time-series data of production equipment operation through edge sensing nodes, construct equipment topology structure according to the process flow dependency relationship between production equipment, and encode the multimodal time-series data based on the equipment topology structure to obtain equipment state representation vector. An anomaly detection unit is used to perform probability distribution modeling on the device state representation vector to generate a state prior distribution and a state posterior distribution, calculate the divergence value between the state prior distribution and the state posterior distribution as the anomaly confidence, and mark the device state representation vector whose anomaly confidence exceeds an adaptive threshold as an anomaly state representation vector. The root cause tracing unit is used to identify related nodes with causal dependencies by using the timestamp corresponding to the abnormal state representation vector as the starting point for tracing, and by using a causal graph inference algorithm to identify related nodes with causal dependencies along the device topology. It extracts the historical device state representation vectors of the related nodes, performs counterfactual reasoning on the historical device state representation vectors to generate comparative state representation vectors, and calculates the causal effect strength between the historical device state representation vectors and the comparative state representation vectors to identify the root cause anomaly source. The early warning control unit is used to construct an early warning propagation path along the device topology based on the root cause anomaly source, generate hierarchical early warning instructions according to the causal effect intensity of each node in the early warning propagation path, and send them to the device control layer for execution.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images