A business hall equipment operation state real-time monitoring method and system

Abstract

The application discloses a kind of business hall equipment operating state real-time monitoring method, it is related to the technical field of electric digital data processing, this method includes: the operating state data of equipment is formed multidimensional time series data by real-time acquisition, with each equipment operating state variable as node, using time-varying granger causality test method to build causal diagram;The time series data after pre-processing and causal diagram are input into causal diagram enhanced prediction transformation network, and the operating state prediction result of equipment in future preset time period is output, compared with threshold value, trigger alarm when abnormal, and calculate quantization attribution score to determine the operating state parameter and contribution percentage that contribute most to prediction anomaly.The network includes dynamic causal diagram learning module, causal intervention module, multiscale causal convolution module and explainability attribution module, and precise causal intervention is realized by directed acyclic graph constraint and stop gradient operation.The application realizes high-precision prediction and quantifiable attribution analysis of the operating state of business hall equipment.

Description

Technical Field

[0001] This application relates to the field of digital data processing technology, and in particular to a method and system for real-time monitoring of the operating status of equipment in a business hall. Background Technology

[0002] As the core location for providing services to customers, the service hall is equipped with a large number of intelligent devices, including self-service terminals, queuing systems, information display screens, network equipment, and servers. The operational status of these devices directly affects the service quality and customer experience of the service hall.

[0003] Currently, several technical solutions exist for monitoring and managing the operational status of service hall equipment. For example, patent CN113157533B determines whether unauthorized operation has occurred by monitoring the open status and movement of the equipment chassis; patent CN114077527A uses middleware to obtain the underlying hardware status information of the equipment and compares it with standard status to achieve anomaly detection; and CN102394785B achieves unified operation and maintenance by performing online detection and fault statistical analysis of service terminal equipment. Furthermore, time series prediction models based on causal reasoning have also been introduced into the field of equipment status prediction, improving the accuracy and interpretability of predictions through causal constraints.

[0004] However, existing technologies still have the following shortcomings: First, existing solutions mainly rely on threshold judgment or rule matching for anomaly detection, lacking the ability to predict the future operating status of equipment and failing to provide early warnings of faults. Second, existing causal reasoning models mainly use backdoor adjustment methods to mitigate the confounding effects of environmental factors, resulting in insufficient robustness in predictions under non-stationary data environments, and their interpretability remains primarily qualitative, lacking the ability to quantitatively attribute prediction results. Third, existing solutions lack the ability to systematically model the relationships between multiple devices within a business hall, making it impossible to assess the cascading risks of a single device's anomaly to the entire equipment network. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this application provides a method and system for real-time monitoring of the operating status of equipment in a business hall, aiming to solve the following technical problems: achieving accurate prediction and early warning of the operating status of equipment in a business hall; improving the predictive robustness of the model in non-stationary data environments through explicit causal mechanism modeling; achieving quantitative attribution analysis of the root causes of equipment failures; and achieving intelligent assessment of the risks of multi-device linkage within the business hall.

[0006] To achieve the above-mentioned objectives, this application provides the following technical solution:

[0007] Firstly, this application provides a method for real-time monitoring of the operating status of equipment in a business hall, including: Step S10: Real-time data collection of equipment operation status is performed by sensors and data acquisition terminals deployed on various equipment in the business hall to form multi-dimensional time series data and perform preprocessing. Furthermore, sensors and data acquisition terminals deployed on various devices in the business hall are used to collect real-time operational status data. This operational status data includes, but is not limited to: CPU utilization, memory usage, network throughput, disk read / write speed, device temperature, runtime, transaction response time, and number of concurrent connections. The collected data is then processed into multi-dimensional time-series data at preset time intervals.

[0008] Step S20: Based on the preprocessed multidimensional time series data, using the operating status variables of each device as nodes, the time-varying Granger causality test method is used to calculate the causal relationship between each variable, and a causal graph with the causal relationship between variables as directed edges is constructed. Furthermore, the collected multidimensional time series data are preprocessed, including missing value imputation, outlier removal, data normalization, and the preprocessed data is sliced ​​according to a preset time window to form training samples and real-time prediction input samples. Based on the preprocessed time series data, the time-varying Granger causality test method is used to calculate the causal relationship between each variable, with the operating status variable of each device as the node, and a causal graph with the causal relationship between variables as directed edges is constructed.

[0009] Specifically, for any two device operating state variables X and Y, a complete vector autoregressive model containing X and a restricted vector autoregressive model without X are established, and the F statistic is calculated. When the F statistic exceeds a preset threshold, X is determined to be a Granger cause of Y, and a directed edge from X to Y is added to the causal graph.

[0010] Step S30: Input the preprocessed multidimensional time series data and the causal graph into the pre-trained causal graph enhancement prediction transformation network for processing, and output the prediction results of the operating status of each device in the future preset time period. Furthermore, the preprocessed time-series data and the causal graph are input into a pre-trained causal graph enhancement prediction transformation network for processing, outputting the predicted operating status of each device within a preset future time period. The causal graph enhancement prediction transformation network includes a dynamic causal graph learning module, a causal intervention module, a multi-scale causal convolution module, and an interpretable attribution module. The dynamic causal graph learning module injects the structural information of the causal graph into the attention mechanism of the prediction network to achieve time-series prediction enhanced by the causal graph. Specifically, the dynamic causal graph learning module fuses the F-statistic of the Granger causality test with the feature interaction weights output by the attention mechanism to generate a time-varying causal matrix. This time-varying causal matrix is ​​then used as a gating mask, applied element-wise to the multi-head attention calculation results. Furthermore, a directed acyclic graph constraint term is used to regularize the causal graph, ensuring that no self-loops or feedback loops appear in the causal graph.

[0011] The causal intervention module is used to input external environmental factors into the network as additional inputs. It models these external environmental factors based on a structural causal model and estimates causal effects through front-door adjustment. These external environmental factors include one or more of business hours, customer traffic levels, and seasonal factors. The causal intervention module introduces a learnable causal structure gating vector and implements causal intervention through gradient stopping operations, ensuring the network learns only through causal paths and avoiding interference from non-causal correlations during network training.

[0012] The multi-scale causal convolution module is used to parallelize and combine dilated causal convolution with a multi-head attention mechanism, simultaneously capturing local temporal patterns and global temporal dependencies. The multi-scale causal convolution module employs a multi-layer stacked structure, with the dilation rate of each layer increasing exponentially to expand the receptive field without increasing the number of parameters.

[0013] The interpretability attribution module is used to calculate the contribution of each input feature to the prediction result using the integral gradient method, and outputs the quantified attribution result.

[0014] The training of the causal graph-enhanced prediction transformation network adopts a multi-task loss function, which is composed of a weighted sum of prediction loss, causal consistency loss and graph structure loss.

[0015] Step S40: Compare the predicted running status with a preset threshold range. If the predicted result exceeds the threshold range, determine that there is an abnormal risk and trigger an alarm. Step S50: The causal graph-enhanced prediction transformation network outputs a quantized attribution score to determine the operating state parameter that contributes the most to the prediction of anomalies and its corresponding contribution percentage.

[0016] Furthermore, the prediction results are compared with a preset threshold range. When the prediction results exceed the threshold range, an abnormal risk is determined and an alarm is triggered. At the same time, the interpretability attribution module of the causal graph-enhanced prediction transformation network outputs a quantified attribution score to determine the operating state parameter that contributes the most to the prediction anomaly and the corresponding contribution percentage.

[0017] The method also includes a multi-device linkage analysis step: based on the causal graph adjacency matrix output by the dynamic causal graph learning module, a device association graph is constructed with each device as a node and the Granger causal relationship between variables as directed edges; and a graph propagation algorithm is used to evaluate the impact of a single device anomaly on other associated devices and the propagation path.

[0018] Secondly, this application also provides a real-time monitoring system for the operating status of business hall equipment, including: Data acquisition and preprocessing module: Used to collect real-time operating status data of equipment through sensors and data acquisition terminals deployed on various equipment in the business hall, form multi-dimensional time series data and perform preprocessing; Cause-effect graph construction module: Based on preprocessed multidimensional time series data, using the operating status variables of each device as nodes, the time-varying Granger causality test method is used to calculate the causal relationship between each variable and construct a cause-effect graph with the causal relationship between variables as directed edges. Causal graph enhancement prediction module: used to input the preprocessed multidimensional time series data and the causal graph into the pre-trained causal graph enhancement prediction transformation network for processing, and output the prediction results of the operating status of each device in the future preset time period; Anomaly detection and alarm module: used to compare the predicted running status with a preset threshold range, and to determine that there is an anomaly risk and trigger an alarm when the predicted result exceeds the threshold range; Anomaly Quantization Attribution Module: Used to output quantization attribution scores through the causal graph-enhanced prediction transformation network to determine the operating state parameters that contribute the most to the prediction of anomalies and their corresponding contribution percentages.

[0019] This application discloses the following technical effects: This application provides a method and system for real-time monitoring of the operating status of equipment in a business hall. Through a dynamic causal graph learning module, the causal structure discovered by Granger causality testing is explicitly injected into the attention mechanism of the prediction network. Furthermore, the causal graph is regularized using directed acyclic graph constraints to ensure its rationality and improve the accuracy of equipment operating status prediction. Compared to existing technologies that implicitly capture the relationships between equipment operating status variables through attention mechanisms, this invention explicitly models the causal relationships between variables, making the prediction results more accurate and reliable.

[0020] Existing anomaly detection methods can only output anomaly results, but cannot explain the cause of the anomaly. This invention uses an integral gradient method through an interpretability attribution module to quantify the percentage contribution of each device's operating status parameter to the predicted anomaly, helping maintenance personnel quickly locate the root cause of the fault.

[0021] Furthermore, the operational data of the business hall equipment is significantly non-stationary due to multiple factors such as business hours, customer traffic, and season. This invention incorporates external environmental factors as additional input into the network through a causal intervention module, explicitly models these factors using a structural causal model, and implements causal intervention through gradient cessation, ensuring the model learns only through causal paths. Compared to existing technologies that use backdoor adjustments to eliminate interference, this invention exhibits superior predictive robustness in non-stationary data environments. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a method for real-time monitoring of the operating status of equipment in a business hall, as provided in an embodiment of this application.

[0023] Figure 2 This is a schematic diagram of the structure of a real-time monitoring system for the operating status of equipment in a business hall, provided as an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Example 1: This application provides a method for real-time monitoring of the operating status of equipment in a business hall, such as... Figure 1 As shown, the method includes: Step S10: Real-time data collection of equipment operation status is performed by sensors and data acquisition terminals deployed on various devices in the business hall, forming multi-dimensional time series data and preprocessing it.

[0026] In this embodiment, a data acquisition terminal is deployed on each smart device in the business hall. These smart devices include self-service terminals, queuing systems, information display screens, network switches, servers, and air conditioning equipment. The data acquisition terminal connects to the business hall equipment via standard interfaces (such as USB, RS-232, or network interfaces).

[0027] The device's operational status data is collected according to a preset sampling period. In this embodiment, the sampling period is preferably once every 30 seconds. The collected operational status data includes, but is not limited to: CPU utilization (%), memory usage (%), network throughput (Mbps), disk read / write speed (MB / s), device temperature (°C), runtime (hours), transaction response time (milliseconds), and number of concurrent connections. The collected operational status data is timestamped and cached locally to ensure that data is not lost in the event of a network interruption.

[0028] The data acquisition terminal transmits the cached operational status data to the edge computing node via the local network of the business hall. This edge computing node, deployed locally in the business hall, preprocesses the received operational status data of the business hall equipment, specifically including: Missing value imputation: Linear interpolation is used to imput missing data points in the running status data; for consecutive missing data segments, a combination of forward imputation and backward imputation is used.

[0029] Outlier removal: A statistical outlier detection method is used to calculate the mean and standard deviation of data for each operating state. Data points that exceed the mean ± 3 times the standard deviation are marked as outliers and removed.

[0030] Data normalization: The minimum-maximum normalization method is used to scale all device operating status data to the [0,1] interval to eliminate the influence between different units.

[0031] After the above preprocessing, standardized multidimensional time series data is obtained, denoted as... ,in Represents a matrix, This represents the total number of time steps. This refers to the number of dimensions of the equipment operation status data, that is, the total number of all operation status parameters collected from all equipment in all business halls.

[0032] Preprocessed standardized data The data is sliced ​​according to a preset time window to form the input samples required for subsequent steps. In this embodiment, the preferred time window width is... =2880 time steps, corresponding to data from the past 24 hours (sampled every 30 seconds, totaling 2880 sampling points over 24 hours); For time t, take the time before that. The running status data of each time step constitutes a time window. : ,in Let be the D-dimensional device operating status data vector collected at time t, and let be the time window. It serves as the data foundation for constructing the causal graph in step S20, and also as the main input to the causal graph enhancement prediction transformation network in step S30.

[0033] After processing in step S10, the raw collected data is converted into standardized time window data that can be used in subsequent steps. This data also preserves the complete values ​​of the operating status variables of each device at all time steps, providing a unified data foundation for the construction of the causal graph in step S20 (which requires analysis of the statistical relationships between variables) and the input of the prediction network in step S30 (which requires providing complete time series features).

[0034] Step S20: Based on the preprocessed multidimensional time series data, using the operating status variables of each device as nodes, the time-varying Granger causality test method is used to calculate the causal relationship between each variable, and a causal graph with the causal relationship between variables as directed edges is constructed.

[0035] This step is based on the standardized time window data output in step S10. Using the operating state variables of each device as nodes, a causal graph between the operating state variables of each device is dynamically constructed using the time-varying Granger causality test method. Specifically, from... Take any two variables from the D dimensions, and perform the following Granger causality test on each pair of variables: For any two device operating state variables X and Y, establish the following two vector autoregressive models, including a complete model that includes X and a restricted model that does not include X:

[0036]

[0037] in, Let Y be the value of the variable at time t. Let X be the value of variable X at time ti. Let be the value of variable Y at time ti, where i is the lag order. and These are the preset regression coefficients; The maximum lag order is determined using the Akaike Information Criterion (AIC) in this embodiment; and This is the residual term.

[0038] Based on the two vector autoregressive models established above, the F-statistic is calculated using the following formula:

[0039] in, This represents the sum of squared residuals in a restricted vector autoregressive model. This is the sum of squared residuals of the complete vector autoregressive model. The sample size is given. When the F-statistic exceeds a preset threshold, X is determined to be a Granger cause of Y, meaning that the historical value of variable X has predictive power for the current value of variable Y. A directed edge from X to Y is then added to the causal graph.

[0040] Since the causal relationships between the operational status variables of the service hall equipment may change over time (for example, the impact of equipment load on response time is more significant during peak business hours), a sliding window strategy is adopted to obtain a time-varying causal matrix. Specifically, for the current time t, the F statistic from the previous w times is taken (w is the width of the sliding window, which is preferably 100 in this embodiment, corresponding to approximately 50 minutes), forming an F statistic sequence. .

[0041] The output of this step is a cause-effect graph, whose adjacency matrix is ​​denoted as . ,element Representing variables Does Granger cause variables? 1 indicates yes, and 0 indicates no. This causal graph will be used as a supervision signal in step S30 to ensure that the causal graph learned by the network is consistent with the Granger causality test results.

[0042] Step S30: Input the preprocessed multidimensional time series data and the causal graph into the pre-trained causal graph enhancement prediction transformation network for processing, and output the prediction results of the operating status of each device in the future preset time period.

[0043] This step will use the standardized time window data output in step S10. The causal graph constructed in step S20 is input into a pre-trained causal graph enhancement prediction transformation network for processing, and the predicted operating status of each device within a preset future time period is output. In this embodiment, the predicted future time period is preferably three scales: 1 hour, 6 hours, and 24 hours, corresponding to 120, 720, and 2880 time steps, respectively.

[0044] The Causal Graph Enhancement Prediction Transformation Network comprises four core modules: a dynamic causal graph learning module, a causal intervention module, a multi-scale causal convolution module, and an interpretable attribution module. The network's data flow is as follows: The causal graph constructed in step S20 is first input into the dynamic causal graph learning module, which outputs the enhanced temporal features of the causal graph; simultaneously The system also inputs the causal intervention module (which is concatenated with external environmental factors) and the multi-scale causal convolution module to generate environmental intervention features and multi-scale temporal features. The outputs of the three modules are concatenated along the channel dimension and then input into the interpretability attribution module to finally output the prediction results.

[0045] The specific structure and function of each module are as follows: The dynamic causal graph learning module first processes the F-statistic sequence obtained from the Granger causality test in step S20. The feature interaction weights output by the attention mechanism built into the dynamic causal graph learning module Fusion to generate a time-varying causal adjacency matrix :

[0046] in, For a moment The adjacency matrix of the time-varying causal graph, whose elements Representing variables For variables The causal strength of is in the range of [0,1], and the larger the value, the stronger the causal relationship. For learnable weight matrix, For splicing operations, Indicates from time At the time This sequence of Granger causality test F-statistics over a continuous time period The attention mechanism built into the dynamic causal graph learning module is based on the input data. The calculated feature interaction weight matrix, This is a normalized exponential function. The technical significance of this fusion operation lies in: merging the results of statistical Granger causality tests... Automatic feature interaction patterns learned by deep learning models Combined, to generate It contains both prior causal knowledge and can adapt to changes in data distribution.

[0047] Next, the dynamic cause-effect graph learning module will... As a gating mask, it is applied to the multi-head attention computation result of the built-in attention mechanism through element-wise multiplication:

[0048] in, The output feature matrix of the dynamic causal graph learning module is the time sequence feature; This represents the standard multi-head attention mechanism. , and These are the query matrix, key matrix, and value matrix, respectively. For dimension expansion operations, in order to perform element-wise multiplication, This is the symbol for element-wise multiplication. The technical significance of this gating mechanism lies in: if If there is no causal relationship between two variables, then even if the original attention mechanism detects a statistical correlation between them, the output of that path will be biased. By masking (multiplying by a value close to 0), the network is made to focus on causal pairs of variables, filtering out spurious correlations at the statistical level.

[0049] The dynamic cause-effect graph learning module also regularizes the cause-effect graph using directed acyclic graph constraints to ensure that the learned cause-effect graph structure is reasonable. ;in, For the constraint values ​​of a directed acyclic graph, when This value is 0 when the graph is a directed acyclic graph; The trace of a matrix is ​​the sum of the elements on its main diagonal. Represents matrix exponentiation. For element-wise multiplication, The total number of operational status variables of the business hall equipment; this constraint term ensures that there are no self-loops and feedback loops in the causal graph by penalizing the existence of directed loops, thereby guaranteeing the directionality and rationality of causal relationships and preventing the network from learning causal structures that are physically impossible to realize.

[0050] The causal intervention module uses external environmental factors as additional input to the network. In this embodiment, external environmental factors include business hours (morning, afternoon, and evening, using 3D one-hot encoding), customer traffic levels (high, medium, and low, using 3D one-hot encoding), and seasonal factors (spring, summer, autumn, and winter, using 4D one-hot encoding), totaling 10 dimensions, denoted as... .

[0051] This module first stores the device operating status data. and external environmental factors coding The features are concatenated along the feature dimension to form a joint feature vector. Then Input is sent to the causal intervention module for processing:

[0052] in, The causal effect estimate output by the causal intervention module has the same dimension as the prediction target; It is a multilayer perceptron network containing two hidden layers (dimensions of 256 and 128, respectively), with ReLU activation function; For a learnable causal structure-gated vector, its elements Indicates the first The probability that each input feature belongs to a causal path; To stop gradient operations, prevent gradients from traversing non-causal paths during backpropagation. Postback. The technical significance of stopping gradient operations is that during model training, parameter updates only update... Network parameters on the path, without updating Parameters on the path. This ensures that the model learns only through causal paths, i.e., only through the gating vector. Only features with larger values ​​(those judged to have a true causal effect) can influence the network's parameter updates, while While the corresponding non-causal features participate in the forward computation, they do not participate in gradient backpropagation, thus avoiding interference from non-causal correlations on network training. Compared to existing techniques that use backdoor adjustments to eliminate profanity, this module achieves explicit modeling of causal paths by stopping gradient operations, making the network model more robust in non-stationary data environments.

[0053] Causal structure gate vector The elements are calculated in the following way: ,in for The One element, The sigmoid activation function compresses the output value to the interval [0,1], making... This can be interpreted as the first The probability that each input feature belongs to a causal path; and These are the learnable weight matrix and bias vector, respectively. For the first The hidden layer representation vector of each input feature in the network layer.

[0054] The multi-scale causal convolution module enhances the network's ability to capture local features and global dependencies in time series data. This module employs a multi-layer stacked structure, with the dilation rate of each layer increasing exponentially.

[0055] in, For the first The dilation rate of a layer dilated causal convolution represents the number of sampling intervals of the convolution kernel on the time axis; The total number of layers is shown in this embodiment. The preferred configuration is 4 layers, with expansion rates of 1, 2, 4, and 8 for each layer. The exponential growth of the expansion rate allows the network to expand its receptive field exponentially without increasing the number of parameters, thus simultaneously capturing both short-term fluctuations in equipment operation (such as sudden load spikes) and long-term trends (such as daily cyclical changes in passenger flow). The calculation process of dilated causal convolution is expressed as follows:

[0056] in, For the first The layer's output at time t, The kernel size is 3 in this embodiment; For the first layer by layer Each convolutional kernel weight, For the first Layer at time Input; is the activation function. Causal convolution ensures that the output at time t depends only on historical data up to and including time t, and does not contain future information, which conforms to the causal principle of time series prediction.

[0057] The final output of the multi-scale causal convolution module is represented as: :

[0058] in, It is a feedforward network containing two linear layers (dimensions 512 and 256 respectively) with ReLU activation function; The output of device operating status data by the multi-head attention mechanism is used to capture global temporal dependencies (such as periodic patterns spanning multiple days). The output of the dilated causal convolution is used to capture local temporal patterns (such as minute-level bursts of anomalies). This parallel combination structure enables the network to focus on both short-term mutations and long-term trends simultaneously, resulting in a faster response to sudden device anomalies.

[0059] The interpretability attribution module uses the integral gradient method to calculate the attribution score of each input feature for the prediction result: ;in, For the network Each input feature in the input sample The integral gradient attribution value under the given conditions, For the current input sample, the first... One eigenvalue; The first input for the baseline The baseline input is a reference benchmark used for comparison with the current input, and adopts a normalized zero vector (i.e., all features take the value of 0, corresponding to the baseline operating state of the business hall equipment). The prediction function of the causal graph enhancement prediction transform network is... For the prediction function on the th The gradient of an input feature represents the sensitivity of that feature to the prediction result; The variable is the integral variable, with a value range of 0 to 1, representing the interpolation path from the baseline input to the current input.

[0060] The normalized attribution score is ,in For the first The normalized attribution score of each input feature, expressed as a percentage, represents the proportion of the feature's contribution to the prediction result. For the first The absolute value of each input feature attribution is used to measure the importance of that feature.

[0061] The output of the dynamic cause-effect graph learning module Output of the causal intervention module and the output of the multi-scale causal convolution module The features are concatenated along the feature dimension to obtain the concatenated features. This feature includes causal enhancement information, environmental intervention information, and multi-scale time-series information; then... The input is fed into a fully connected layer for the final prediction output, yielding the prediction result. ,in The prediction time steps (120, 720, or 2880 depending on the prediction scale).

[0062] The causal graph augmentation prediction transformation network needs to be pre-trained before it can be used. The specific training process is as follows: Collect historical operating data of equipment in the business hall for at least one year, including time series data of operating status variables of each equipment and corresponding records of external environmental factors (business hours, customer flow level, seasonal factors, etc.); after the same preprocessing as in step S10, training dataset, validation dataset and test dataset are formed, and the preferred ratio of the three is 7:1.5:1.5.

[0063] Training samples are constructed from the training dataset using a sliding window approach. Each training sample consists of an input part and a label part: the input part includes a window of historical time series data. (Window width) =2880 time steps), based on this historical time window, the causal graph extracted by the method described in step S20, and the external environmental factor encoding vector within the corresponding time period. The tagged section represents the future after a historical time window. Actual equipment operating status data at each time step In this embodiment, three models are trained to correspond to three prediction scales: 120 (next 1 hour), 720 (next 6 hours), and 2880 (next 24 hours).

[0064] The input portion of the training samples is fed into the causal graph augmented prediction transform network, and the network outputs the future... Predicted equipment operating status at each time step The predicted value With tag value The comparison is performed, and the loss value is calculated using the multi-task loss function, which is defined as follows: ,in To predict the loss, mean squared error is used to measure the difference between the predicted value and the label value; The causal consistency loss is used to constrain the causal graph learned by the network. Causal graph extracted with Granger causality test Maintain consistency The graph structure loss, used to constrain the directed acyclicity of the causal graph, is represented as: , and The balance coefficient is preferably 0.1 and 0.05 in this embodiment.

[0065] The backpropagation algorithm is used to calculate the gradient of the loss value with respect to each network parameter, and the network parameters are updated using an optimizer. Training employs the Adam optimizer with an initial learning rate of 0.001, using a cosine annealing learning rate scheduling strategy. The batch size is 64, and the maximum number of training epochs is 200. After each training epoch, the model performance is evaluated on the validation set. Training is terminated early when the validation set loss no longer decreases for 20 consecutive epochs to prevent overfitting. After training, the model with the minimum loss on the validation set is selected as the final model, and its final performance is evaluated on the test set.

[0066] The trained causal graph augmentation prediction transformation network is deployed on a cloud server for real-time inference. During inference, the network parameters are fixed, and the input is the currently collected time window data and the causal graph extracted in real time in step S20. The output is the prediction results of the operating status of each device. The network is retrained every 30 days or after accumulating enough new data to adapt to changes in device operating modes.

[0067] Step S40: Compare the predicted running status with a preset threshold range. If the predicted result exceeds the threshold range, it is determined that there is an abnormal risk and an alarm is triggered.

[0068] In this embodiment, the predicted device operating status output in step S30 is... The system compares the predicted value with a preset normal operating status threshold range to determine if any abnormal risks exist. This threshold range is determined based on equipment specifications and historical normal operating data. When the predicted result of any operating status variable exceeds its corresponding threshold range, the system determines that the equipment has an abnormal operating risk and generates an alarm message. The alarm message includes: equipment identifier, anomaly type (such as excessively high CPU utilization or abnormal response time), predicted occurrence time, and predicted severity (i.e., the standard deviation multiple of the deviation from the mean).

[0069] Step S50: The causal graph-enhanced prediction transformation network outputs a quantized attribution score to determine the operating state parameter that contributes the most to the prediction of anomalies and its corresponding contribution percentage.

[0070] In this embodiment, when an alarm is triggered in step S40, the interpretability attribution module in the causal graph-enhanced prediction transformation network outputs a quantized attribution score to determine the operating state parameter that contributes most to the prediction anomaly and its corresponding contribution percentage. Specifically, the prediction sample that triggered the alarm is input into the interpretability attribution module, and the integral gradient method described in step S30 is used to calculate the attribution score of each input feature to the prediction result. The data is then normalized according to the normalization formula described in step S30, and the contribution ratio of each feature to the predicted anomaly is output as a percentage. The technical significance of this quantification attribution process is that by calculating the gradient on the integral path from the baseline to the current input (the sample that triggered the alarm), the final prediction result is decomposed into the sum of the marginal contributions of each input feature, thereby accurately locating the key factors that lead to the predicted anomaly.

[0071] Based on the quantitative attribution report, operations and maintenance personnel can quickly identify the main factors causing abnormal response times and then take targeted measures such as rate limiting or capacity expansion, without having to investigate all possible causes of failure one by one.

[0072] In addition, the method also includes multi-device linkage analysis: based on the causal graph output by the dynamic causal graph learning module in the causal graph enhanced prediction transformation network, a device association graph is constructed with each business hall device as a node and the Granger causal relationship between variables as directed edges. Based on the device association diagram and the prediction results of devices in each business hall, a graph propagation algorithm is used to evaluate the impact of a single device malfunction on other associated devices and the propagation path: This can be expressed by the formula:

[0073] in, This is the anomaly propagation probability matrix, and its elements are... Indicates device The abnormal propagation to the device The probability of; The softmax function is row-normalized, and the adjacency matrix of the time-varying causal graph is... Each row is converted into a probability distribution form; This is an abnormal status vector for each business hall's equipment, where each element represents the severity of the abnormality for each piece of equipment.

[0074] The real-time monitoring method for the operating status of business hall equipment provided in this embodiment, through the complete process of steps S10 to S50, achieves end-to-end intelligent monitoring of the operating status of business hall equipment from perception to prediction to attribution. This is achieved through data acquisition and preprocessing, causal graph construction, causal graph enhancement prediction transformation network processing, anomaly detection, and quantitative attribution. Its core improvements are reflected in: integrating Granger causality testing with attention mechanisms through a dynamic causal graph learning module and applying DAG constraints; explicitly modeling causal paths using stopping gradient operations through a causal intervention module; capturing local and global temporal dependencies in parallel through a multi-scale causal convolution module; and achieving quantitative attribution of predicted anomalies through an integral gradient method. These improvements collectively enhance the accuracy of equipment operating status prediction and robustness in non-stationary data environments, effectively improving the intelligent level of business hall equipment operation and maintenance.

[0075] Example 2: The real-time monitoring system for the operating status of business hall equipment provided in this embodiment of the invention can execute the real-time monitoring method for the operating status of business hall equipment provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method, such as... Figure 2 As shown, it includes: Data acquisition and preprocessing module: Used to collect real-time operating status data of equipment through sensors and data acquisition terminals deployed on various equipment in the business hall, form multi-dimensional time series data and perform preprocessing; Cause-effect graph construction module: Based on preprocessed multidimensional time series data, using the operating status variables of each device as nodes, the time-varying Granger causality test method is used to calculate the causal relationship between each variable and construct a cause-effect graph with the causal relationship between variables as directed edges. Causal graph enhancement prediction module: used to input the preprocessed multidimensional time series data and the causal graph into the pre-trained causal graph enhancement prediction transformation network for processing, and output the prediction results of the operating status of each device in the future preset time period; Anomaly detection and alarm module: used to compare the predicted running status with a preset threshold range, and to determine that there is an anomaly risk and trigger an alarm when the predicted result exceeds the threshold range; Anomaly Quantization Attribution Module: Used to output quantization attribution scores through the causal graph-enhanced prediction transformation network to determine the operating state parameters that contribute the most to the prediction of anomalies and their corresponding contribution percentages.

[0076] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0077] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for real-time monitoring of the operating status of equipment in a business hall, characterized in that, The method includes: Step S10: Real-time data collection of equipment operation status is performed by sensors and data acquisition terminals deployed on various equipment in the business hall to form multi-dimensional time series data and perform preprocessing. Step S20: Based on the preprocessed multidimensional time series data, using the operating status variables of each device as nodes, the time-varying Granger causality test method is used to calculate the causal relationship between each variable, and a causal graph with the causal relationship between variables as directed edges is constructed. Step S30: Input the preprocessed multidimensional time series data and the causal graph into the pre-trained causal graph enhancement prediction transformation network for processing, and output the prediction results of the operating status of each device in the future preset time period: The causal graph enhancement prediction transformation network includes a dynamic causal graph learning module, a causal intervention module, a multi-scale causal convolution module, and an interpretable attribution module. The dynamic causal graph learning module is used to inject the structural information of the causal graph into the attention mechanism of the prediction network to achieve causal graph-enhanced temporal prediction. Based on the causal graph output by the dynamic causal graph learning module in the causal graph augmented prediction transformation network, a device association graph is constructed with each business hall device as a node and the Granger causal relationship between variables as directed edges. The causal intervention module is used to access the network as an additional input of external environmental factors, model the external environmental factors based on the structural causal model, and estimate the causal effect through front door adjustment. The external environmental factors include business hours, customer flow level and seasonal factors. The causal intervention module introduces a learnable causal structure matrix and achieves causal intervention by stopping gradient operations. This process is expressed by the following formula: in, The causal effect estimate output by the causal intervention module. It is a multilayer perceptron network. The input feature vector for this module is formed by concatenating the encoded vectors of device operating status data and external environmental factors. For a learnable causal structure-gated vector, its elements Indicates the first The probability that each input feature belongs to a causal path; For element-wise multiplication, To stop gradient operations, prevent gradients from traversing non-causal paths during backpropagation. Backhaul ensures the network only uses causal paths. To learn and avoid interference from non-causal correlations in network training; The multi-scale causal convolution module is used to combine dilated causal convolution with multi-head attention mechanism in parallel, while capturing local temporal patterns and global temporal dependencies. The interpretable attribution module is used to calculate the contribution of each input feature to the prediction result using the integral gradient method, and outputs the quantified attribution result. Step S40: Compare the predicted running status with a preset threshold range. If the predicted result exceeds the threshold range, determine that there is an abnormal risk and trigger an alarm. Step S50: The quantized attribution score is output through the causal graph enhanced prediction transformation network to determine the running state parameter that contributes the most to the prediction anomaly and the corresponding contribution percentage. Based on the device association diagram and the prediction results of devices in each business hall, a graph propagation algorithm is used to evaluate the impact of a single device malfunction on other associated devices and the propagation path: This can be expressed by the formula: in, This is the anomaly propagation probability matrix, and its elements are... Indicates equipment Abnormal propagation to the device The probability of; The softmax function is row-normalized, and the adjacency matrix of the time-varying causal graph is... Each row is converted into a probability distribution form; This is an abnormal status vector for each business hall's equipment, where each element represents the severity of the abnormality for each piece of equipment.

2. The method for real-time monitoring of the operating status of business hall equipment as described in claim 1, characterized in that, In step S20, the time-varying Granger causality test method determines the causal relationship for any two operational state variables X and Y of the business hall equipment in the following way: Construct a complete vector autoregressive model including X and a restricted vector autoregressive model without X, and calculate the F-statistic, expressed by the following formula: in, This represents the sum of squared residuals in a restricted vector autoregressive model. This is the sum of squared residuals of the complete vector autoregressive model. The maximum lag order, The sample size is given. When the F-statistic exceeds a preset threshold, X is determined to be a Granger cause of Y, meaning that the historical value of variable X has predictive power for the current value of variable Y. A directed edge from X to Y is then added to the causal graph.

3. The method for real-time monitoring of the operating status of business hall equipment as described in claim 1, characterized in that, The dynamic causal graph learning module fuses the F-statistic of the time-varying Granger causality test with the feature interaction weights output by the attention mechanism to generate a time-varying causal graph adjacency matrix, expressed by the formula: in, For a moment The adjacency matrix of the time-varying causal graph. For learnable weight matrix, For splicing operations, Indicates from time At the time This sequence of Granger causality test F-statistics over a continuous time period This refers to the feature interaction weight matrix output by the attention mechanism built into the dynamic causal graph learning module. It is a normalized exponential function.

4. The method for real-time monitoring of the operating status of business hall equipment as described in claim 3, characterized in that, The dynamic cause-effect graph learning module will also As a gating mask, it is applied to the multi-head attention calculation result of the built-in attention mechanism through element-wise multiplication, and the causal graph is regularized by a directed acyclic graph constraint term, expressed by the following formula: in, For the constraint values ​​of a directed acyclic graph, when This value is 0 when the graph is a directed acyclic graph; The trace of a matrix is ​​the sum of the elements on its main diagonal. Represents matrix exponentiation. For element-wise multiplication, The number of operational status variables of the business hall equipment; this constraint ensures that no self-loops and feedback loops appear in the causal graph by penalizing the existence of directed cycles.

5. The method for real-time monitoring of the operating status of business hall equipment as described in claim 1, characterized in that, The multi-scale causal convolution module adopts a multi-layer stacked structure, and the expansion rate of each layer's convolution kernel increases exponentially to expand the receptive field without increasing the number of parameters, while capturing temporal dependency patterns at different time scales. This module also concatenates the output of dilated causal convolution with the output of multi-head attention mechanism in the channel dimension and then fuses them through a feedforward network to achieve parallel capture of local temporal patterns and global temporal dependencies.

6. The method for real-time monitoring of the operating status of business hall equipment as described in claim 1, characterized in that, The interpretable attribution module uses the integral gradient method to calculate the attribution score of each input feature to the prediction result, expressed by the formula: in, For the network Each input feature in the input sample The integral gradient attribution value under the given conditions, For the current input sample, the first... One eigenvalue; The first input for the baseline The baseline input is a reference benchmark used for comparison with the current input, and is a normalized zero vector; The prediction function of the causal graph enhancement prediction transform network is... For the prediction function on the th The gradient of an input feature represents the sensitivity of that feature to the prediction result; The variable is the integral variable, with a value range of 0 to 1, representing the interpolation path from the baseline input to the current input.

7. A real-time monitoring system for the operating status of equipment in a business hall, characterized in that, The system is used to implement the real-time monitoring method for the operating status of business hall equipment as described in any one of claims 1-6, and the system includes: Data acquisition and preprocessing module: Used to collect real-time operating status data of equipment through sensors and data acquisition terminals deployed on various equipment in the business hall, form multi-dimensional time series data and perform preprocessing; Cause-effect graph construction module: Based on preprocessed multidimensional time series data, using the operating status variables of each device as nodes, the time-varying Granger causality test method is used to calculate the causal relationship between each variable and construct a cause-effect graph with the causal relationship between variables as directed edges. Causal graph enhancement prediction module: used to input the preprocessed multidimensional time series data and the causal graph into the pre-trained causal graph enhancement prediction transformation network for processing, and output the prediction results of the operating status of each device in the future preset time period; Anomaly detection and alarm module: used to compare the predicted running status with a preset threshold range, and to determine that there is an anomaly risk and trigger an alarm when the predicted result exceeds the threshold range; Anomaly Quantization Attribution Module: Used to output quantization attribution scores through the causal graph-enhanced prediction transformation network to determine the operating state parameters that contribute the most to the prediction of anomalies and their corresponding contribution percentages.

Images