Factory data anomaly detection method and system based on big data analysis

Abstract

The application relates to the field of data processing, in particular to a factory data anomaly detection method and system based on big data analysis. The method comprises the following steps: calculating the change trend of each dimension working parameter at each moment, clustering according to the parameter value and the change trend at each moment, using the cluster number of the cluster after clustering to replace the features at each moment to obtain a transformed sequence, setting a hyperparameter pair, calculating the transfer entropy of the transformed sequence and a label sequence, and then obtaining a stable gain, obtaining the best input length of each dimension working parameter according to the stable gain, constructing an anomaly detection model of each dimension, and weighting and summing the detection results of each dimension according to the accuracy of the model to obtain the final detection result. The application determines the best input length by finding the relatively stable condition between each dimension parameter and the label, and has the effect of improving the prediction accuracy of the model.

Description

Technical Field

[0001] This application relates to the field of data processing, and in particular to a method and system for detecting anomalies in factory data based on big data analysis. Background Technology

[0002] With the advancement of industrialization and intelligent manufacturing, manufacturing production lines have accumulated a large amount of industrial time-series data through sensors and other equipment. This data provides important information for system operation status control, fault diagnosis, and early warning. However, manufacturing systems suffer from problems such as product quality defects, equipment failures, and performance degradation. Anomaly detection, fault monitoring, and equipment health status analysis are crucial tasks for achieving lean production and intelligent manufacturing. Traditional anomaly detection methods (such as statistical methods and rule-based thresholds) often struggle to meet the requirements of accuracy, efficiency, and reliability when dealing with high-dimensional, multi-source, heterogeneous, and highly dynamic industrial big data.

[0003] Factory data anomaly detection methods based on big data analytics can cope with the complexity and scale of industrial big data. By learning from historical data through technologies such as machine learning and deep learning, normal behavior patterns can be established, thereby accurately identifying abnormal data, realizing fault early warning and diagnosis, and ensuring the safety and efficiency of industrial production.

[0004] The patent with publication number CN117556366A discloses a data anomaly detection system and method based on data filtering. The method involves setting a distance value k and a distance influence coefficient to perform multiple anomaly detections on all device data points, obtaining several suspected abnormal data points and the anomaly factor corresponding to each suspected abnormal data point in each anomaly detection, and filtering all suspected abnormal data points according to the degree of anomaly to obtain all abnormal data points.

[0005] This method only considers the characteristics of data at a single point in time to calculate the anomaly degree, without taking into account the continuous characteristics of factory data, and also without considering the importance of different characteristics for anomaly detection. Summary of the Invention

[0006] To address the problem that existing factory data anomaly detection methods do not consider the continuous nature of factory data and the importance of different features for anomaly detection, resulting in low anomaly detection accuracy, this application provides a factory data anomaly detection method and system based on big data analysis.

[0007] Firstly, this application provides a factory data anomaly detection method based on big data analysis, employing the following technical solution:

[0008] The factory data anomaly detection method based on big data analysis includes the following steps: constructing a multi-dimensional working parameter sequence of factory equipment and a corresponding anomaly label sequence; for any dimension of the working parameter sequence, calculating the change trend at each time moment; clustering the parameter values ​​and change trends at each time moment to obtain multiple clusters; numbering the clusters, using the numbers as variable values ​​at each time moment within the cluster; and constructing a transformation sequence based on the variable values.

[0009] Set an initial hyperparameter pair, calculate the transfer entropy between the downtransformed sequence of the hyperparameter pair and the anomaly label sequence, calculate the stable gain based on the transfer entropy, traverse the hyperparameter pairs, and obtain the optimal hyperparameter pair when the stable gain reaches its maximum.

[0010] When the two hyperparameter values ​​in the optimal hyperparameter pair are equal, the value of either hyperparameter is the optimal input length of the anomaly detection model; otherwise, the value of the largest hyperparameter in the optimal hyperparameter pair is the optimal input length of the anomaly detection model, and the anomaly detection model is constructed based on the optimal input length.

[0011] The above operations are performed on the working parameter sequences of all dimensions to obtain the anomaly detection model for each dimension. The anomaly detection results of each dimension anomaly detection model are weighted and summed to obtain the degree of anomaly of the factory data.

[0012] The beneficial effects are as follows: By calculating the changing trend of the parameter sequence at each time step, the characteristics of the parameter at the current time step and the characteristics of the parameter changing over time are obtained. The characteristics of the parameters at each time step are clustered. After clustering, the cluster numbers are used to replace the characteristics of the parameters at each time step. Then, the transfer entropy of different hyperparameter pairs for each parameter sequence and label is calculated. Based on the transfer entropy, the stable gain is calculated, and the optimal hyperparameter pair is selected. Anomaly detection models are constructed for each parameter sequence based on the size of the hyperparameter pair. The detection results of different anomaly detection models are weighted and summed to obtain the final detection result. By analyzing the relationship between parameters and labels, the optimal input length for each parameter anomaly detection model is obtained. By weighting the detection results of different parameter anomaly result models according to the relationship between parameters and labels, the detection accuracy of the model is improved.

[0013] Optionally, the multi-dimensional operating parameter sequence includes a temperature sequence, a current sequence, and a voltage sequence.

[0014] Optionally, the method for calculating the trend of change is as follows: set a time window, and take the result of the trend test of the parameter value sequence within the time window as the trend of change.

[0015] The beneficial effects are: trend testing can obtain the probability that the current parameter value has an upward trend, a stable trend or a downward trend, and trend testing can obtain the characteristics of the change of working parameters over time at each moment.

[0016] Optionally, the method for calculating the trend of change is as follows: set a time window and use the variance of the parameter value sequence within the window as the trend of change.

[0017] The beneficial effect is that variance reflects the characteristics of the change of working parameters over time at each moment; the larger the variance, the greater the fluctuation; the smaller the variance, the smaller the fluctuation.

[0018] Optionally, the distance calculation method during clustering is as follows: the parameter values ​​and changing trends of the working parameter sequence at the same time are combined into a vector, and the Euclidean distance between vectors at different times is the distance during clustering.

[0019] The beneficial effects are: clustering the features of the working parameter sequence at different times can group similar working parameter features into one class and dissimilar working parameter features into different classes. Using the clustering results to reflect the features of the working parameters at each time point reduces the amount of computation in subsequent steps.

[0020] Optionally, the hyperparameter pair consists of transform sequence hyperparameters and anomaly label sequence hyperparameters.

[0021] The beneficial effects are as follows: the value of the hyperparameter is used to calculate the transfer entropy of the transformed sequence to the anomalous label sequence. The transfer entropy can reflect the reduction of uncertainty of the transformed sequence to the anomalous label sequence. The greater the reduction of uncertainty, the more important the transformation sequence is in predicting the anomalous label.

[0022] Optionally, the stabilizing gain is the ratio of the transfer entropy of the transformed sequence to the anomalous label sequence to the conditional entropy of the anomalous label sequence.

[0023] The beneficial effects are: the stabilization gain takes into account the instability of the abnormal label sequence itself, while considering the increase of label stability by the transformation sequence, and determines the input length of the anomaly detection model in each dimension through the stabilization gain.

[0024] Secondly, this application provides a factory data anomaly detection system based on big data analysis, employing the following technical solution:

[0025] The factory data anomaly detection system based on big data analysis includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the factory data anomaly detection method based on big data analysis described above.

[0026] The beneficial effect is that the above-mentioned factory data anomaly detection method based on big data analysis is generated into a computer program and stored in the memory so that it can be loaded and executed by the processor. Thus, the system can be made according to the memory and processor, which is convenient to use.

[0027] This application has the following technical effects:

[0028] 1. By clustering the parameter values ​​and trends of the working parameter sequence at each time point, the features of the working parameter sequence at each time point are replaced by the cluster number, so that the features of each time point of the replaced sequence can reflect both static features and dynamic changes.

[0029] 2. The optimal input length of the model is determined by the stable gain. Under the optimal input length, the relationship between the input working parameter sequence and the output label is relatively certain and the stability is higher. Therefore, the model is more likely to learn the relationship between the two, and the model performs better, thus improving the prediction accuracy of the model. Attached Figure Description

[0030] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts.

[0031] Figure 1 This is a flowchart illustrating steps S1-S4 of the factory data anomaly detection method based on big data analysis in this application.

[0032] Figure 2 This is a structural block diagram of the factory data anomaly detection system based on big data analysis proposed in this application. Detailed Implementation

[0033] This application discloses a factory data anomaly detection method based on big data analysis, referring to... Figure 1 This includes the following steps:

[0034] S1: Construct a multi-dimensional sequence of working parameters for factory equipment and a corresponding sequence of anomaly labels.

[0035] In one embodiment, a temperature sensor is used to collect the temperature value of the device, a current sensor is used to collect the current value of the device, and a voltage sensor is used to collect the voltage value of the device. Those skilled in the art assign a label to each collected data according to the actual scenario, and the label is normal / abnormal / potentially risky. The working parameter sequence is normalized to eliminate the influence of dimensions.

[0036] Thus, we have obtained the multi-dimensional operating parameter sequence of the factory equipment and the corresponding anomaly label sequence.

[0037] S2: For any dimension of working parameter sequence, calculate the change trend at each time point, cluster the parameter values ​​and change trends at each time point to obtain multiple clusters, number the clusters, and use the number as the variable value at each time point within the cluster, and construct the transformation sequence based on the variable value.

[0038] In one embodiment, in factory data anomaly detection, data anomalies are caused by two main reasons: firstly, parameter values ​​are too large, exceeding the normal operating range of the equipment; secondly, changes or fluctuations in parameters affect the normal operation of the equipment. Therefore, the trend of change in the working parameter sequence at each moment is calculated. Specifically, for any dimension of the working parameter sequence, the calculation process of the trend is as follows: for any moment, a preset time window is set, and the working parameter sequence within the window is subjected to trend testing. The trend testing method is the Mann-Kendall trend test. After the trend test, the probability of the working parameter sequence having an upward trend, a downward trend, or a stable trend is obtained. This process is repeated for each moment of the working parameters to obtain the trend of change at each moment.

[0039] In another embodiment, the calculation process of the trend of change is as follows: for any given moment, a preset time window is set, and the variance of the sequence parameter values ​​within the window is calculated. The variance reflects the fluctuation of the parameter values. The larger the variance, the more drastic the change of the parameter values ​​at that moment. The smaller the variance, the smoother the change of the parameter values ​​at that moment.

[0040] After trend testing, the characteristics of each time step in the working parameter sequence include parameter values ​​and trends. Clustering is then performed on these characteristics, and the distance between clusters is calculated using the following formula:

[0041]

[0042] in, This represents the distance between the features at time i and time j of the working parameter sequence. This represents the parameter value at time i in the working parameter sequence. This represents the parameter value at time j in the working parameter sequence. This represents the trend of change in the working parameter sequence at time i. This represents the trend of change in the working parameter sequence at time j.

[0043] The features of each time step in the working parameter sequence are clustered using the distance calculation formula described above. The clustering method can be k-means clustering or density-based clustering. After clustering, the clusters are numbered, and the cluster numbers are used to replace the features of each time step within the cluster to obtain the transformed sequence. For example, a working parameter sequence of length 10 is clustered, resulting in 3 clusters. These 3 clusters are then numbered. After numbering, cluster 1 contains features from time steps 1, 4, and 6; cluster 2 contains features from time steps 3, 7, and 8; and cluster 3 contains features from time steps 2, 5, 9, and 10. The resulting transformed sequence is: .

[0044] Perform the above operation on the working parameter sequence for each dimension to obtain the transformation sequence for each dimension.

[0045] S3: Set initial hyperparameter pairs, calculate the transfer entropy between the downtransformed sequence of the hyperparameter pair and the anomaly label sequence, calculate the stable gain based on the transfer entropy, traverse the hyperparameter pairs, and obtain the optimal hyperparameter pair when the stable gain reaches its maximum. When the two hyperparameter values ​​in the optimal hyperparameter pair are equal, any one of the hyperparameter values ​​is the optimal input length of the anomaly detection model; otherwise, the largest hyperparameter value in the optimal hyperparameter pair is the optimal input length of the anomaly detection model. Construct the anomaly detection model based on the optimal input length.

[0046] In one embodiment, to obtain an accurate predictive model for factory data anomaly detection, the input length of the model needs to be considered. This is because the complexity of the relationship between the working parameter sequence and anomaly detection varies with different input lengths. When the complexity of the relationship between the input parameter sequence and anomaly labels is low, the model is more likely to learn the mapping relationship between the two. Therefore, it is necessary to calculate the optimal input length of the working parameters. Specifically:

[0047] For the target dimension's transform sequence and label sequence, a set of hyperparameter pairs is preset. Each hyperparameter pair consists of a set of hyperparameters, including the hyperparameters of the transform sequence and the hyperparameters of the label sequence. The transfer entropy between the transform sequence and the label sequence under the hyperparameter pair is calculated using the following formula:

[0048]

[0049] in, Let X represent the transfer entropy of the transformation sequence X to the label sequence Y under hyperparameters m and n. This represents the labels of the label sequence Y at times m before time t. This represents the label of the label sequence Y at time t+1. Let Y represent the conditional entropy of the label sequence Y at the next time step, given the labels at the previous m time steps. This represents the parameter values ​​of the transformed sequence X at n time points prior to time t. Let Y represent the label values ​​of the known label sequence Y at m time steps before time t, the parameter values ​​of the transformed sequence X at n time steps before time t, and the conditional entropy of the label sequence Y at the next time step.

[0050] Transfer entropy reflects the increase in label stability after introducing a transformation sequence. The first term in transfer entropy represents the stability of future labels given historical labels, while the second term represents the stability of future labels given historical labels and parameters. Therefore, the larger the value of transfer entropy, the better the parameter's predictive effect on the label. When the value of transfer entropy is 0, it means that the parameter has no effect on the label's prediction because it does not reduce the label's instability. When the value of the second term of transfer entropy is 0, the transfer entropy reaches its maximum, indicating that the future of the label is completely predicted after introducing the transformation sequence.

[0051] A higher transfer entropy indicates a greater reduction in uncertainty of the transformed sequence relative to the label sequence under that hyperparameter pair. Since the initial uncertainty of the label differs under different hyperparameters, to quantify the reduction in uncertainty of the transformed sequence pair, the stability gain of the transformed sequence relative to the label sequence under the hyperparameter pair is calculated. The specific calculation formula is as follows:

[0052]

[0053] in, This represents the stable gain of the transformation sequence X on the label sequence Y under hyperparameters m and n. Let X represent the transfer entropy of the transformation sequence X to the label sequence Y under hyperparameters m and n. This represents the labels in the label sequence Y up to time m before time t. This represents the label at time t+1 in the label sequence Y. Let Y represent the conditional entropy of the label sequence Y at the next time step, given the labels at the previous m time steps.

[0054] A larger transfer entropy results in a larger stability gain, indicating a greater increase in the determinism of the label sequence due to the transformation sequence, thus leading to a larger stability gain. Conversely, a smaller conditional entropy of the label indicates a smaller uncertainty in the label sequence, also resulting in a larger stability gain. The stability gain is calculated iteratively for different hyperparameter pairs. The hyperparameter pair corresponding to the maximum stability gain is considered the optimal pair. The optimal input length for the target dimension anomaly detection model is determined based on the values ​​of the two hyperparameters in the pair. Specifically:

[0055] When the hyperparameter values ​​of the transform sequence and the label sequence are equal in the optimal hyperparameter pair, any one of the hyperparameter values ​​is the optimal input length of the anomaly detection model; when the hyperparameter values ​​of the transform sequence and the label sequence are not equal in the optimal hyperparameter pair, the larger hyperparameter value is the optimal input length of the anomaly detection model, and the anomaly detection model is constructed based on the optimal input length.

[0056] S4: Perform the above operations on the working parameter sequences of all dimensions to obtain the anomaly detection model for each dimension. Then, sum the detection results of the anomaly detection models for each dimension in a weighted manner to obtain the degree of anomaly in the factory data.

[0057] In one embodiment, the dataset for training the anomaly detection model is constructed based on the optimal input length of the target dimension. When the hyperparameter values ​​of the target dimension transformation sequence and the label sequence are equal, the input to the anomaly detection model is the target dimension sequence value and the label sequence value with the optimal input length, and the output is the label value at the next time step. When the hyperparameter values ​​of the target dimension transformation sequence and the label sequence are not equal, to ensure that the input length of the target dimension is the same as the input length of the label sequence, the sequence with the smaller hyperparameter is padded with zeros. For example, the hyperparameter value of the target dimension transformation sequence is 4, and the hyperparameter value of the label sequence is 3. Then, the target dimension sequence input each time is... The label sequence is The output is the label value at the next time step. After the training dataset for the target dimension is constructed, the anomaly detection model for the target dimension is trained using this dataset. The model's input is the target dimension sequence and the label sequence, and the model's output is the label value at the next time step. The model's loss function is cross-entropy loss, and the model parameters are updated using the gradient descent algorithm. The model is trained multiple times, and the optimal model is selected using the model's accuracy as an evaluation metric.

[0058] The above operations are performed on each dimension to obtain the optimal model for each dimension. In real-time anomaly detection, multi-dimensional working parameter sequences of factory equipment are collected, and the data of each dimension are input into the anomaly detection model of each dimension. Based on the accuracy of the anomaly detection model of each dimension, the detection results of the anomaly detection model of each dimension are weighted and summed together to obtain the anomaly degree of the final factory data.

[0059] This application also discloses a factory data anomaly detection system based on big data analysis, such as... Figure 2 As shown, it includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the factory data anomaly detection method based on big data analysis according to this application.

[0060] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0061] In this application, the aforementioned memory can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store required information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device.

[0062] While this specification has shown and described numerous embodiments of the present application, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of the present application. It should be understood that various alternatives to the embodiments of the present application described herein may be employed in the practice of this application.

[0063] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A factory data anomaly detection method based on big data analysis, characterized in that, Includes the following steps: Construct a multi-dimensional sequence of working parameters and corresponding anomaly label sequences for factory equipment. For any dimension of the working parameter sequence, calculate the trend of change at each time point. Cluster the parameter values ​​and trends at each time point to obtain multiple clusters. Number the clusters and use the numbers as variable values ​​at each time point within the cluster. Construct a transformation sequence based on the variable values. Set an initial hyperparameter pair, which is a set of hyperparameters including the hyperparameters of the transform sequence and the hyperparameters of the label sequence. Calculate the transfer entropy between the transform sequence and the anomalous label sequence under the hyperparameter pair. Calculate the stabilization gain based on the transfer entropy. Iterate through the hyperparameter pairs and obtain the optimal hyperparameter pair when the stabilization gain reaches its maximum. The stabilization gain is the ratio of the transfer entropy of the transform sequence to the anomalous label sequence to the conditional entropy of the anomalous label sequence. When the two hyperparameter values ​​in the optimal hyperparameter pair are equal, the value of either hyperparameter is the optimal input length of the anomaly detection model; otherwise, the value of the largest hyperparameter in the optimal hyperparameter pair is the optimal input length of the anomaly detection model, and the anomaly detection model is constructed based on the optimal input length. The above operations are performed on the working parameter sequences of all dimensions to obtain the anomaly detection model for each dimension. The anomaly detection results of each dimension anomaly detection model are weighted and summed to obtain the degree of anomaly of the factory data.

2. The factory data anomaly detection method based on big data analysis according to claim 1, characterized in that, The multi-dimensional operating parameter sequence includes a temperature sequence, a current sequence, and a voltage sequence.

3. The factory data anomaly detection method based on big data analysis according to claim 1, characterized in that, The method for calculating the trend is as follows: set a time window, and take the result of the trend test of the parameter value sequence within the time window as the trend.

4. The factory data anomaly detection method based on big data analysis according to claim 1, characterized in that, The method for calculating the trend of change is as follows: set a time window and use the variance of the parameter value sequence within the window as the trend of change.

5. The factory data anomaly detection method based on big data analysis according to claim 1, characterized in that, The distance calculation method during clustering is as follows: the parameter values ​​and changing trends of the working parameter sequence at the same time are combined into a vector, and the Euclidean distance between vectors at different times is the distance during clustering.

6. A factory data anomaly detection system based on big data analysis, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the factory data anomaly detection method based on big data analysis according to any one of claims 1-5.

Images