Corrugated board production process anomaly monitoring method and system

Abstract

The present application relates to the technical field of anomaly monitoring, and particularly relates to a corrugated board production process anomaly monitoring method and system. The method comprises the following steps: obtaining a historical normal data set and a real-time data stream, constructing a reference Gaussian mixture model for different production brand conditions, obtaining a reference statistical moment vector, determining a fusion weight based on the reference statistical moment vector and a statistical moment vector of the real-time data stream, calculating a weight entropy and a composite reference probability density function, adaptively selecting an analysis window according to a kurtosis value of the real-time data stream, constructing a current probability density function and calculating an asymmetric penalty relative entropy between the current probability density function and the composite reference probability density function, selecting a judgment threshold according to a current device running speed and an environmental humidity, and determining an anomaly when the asymmetric penalty relative entropy is greater than the selected judgment threshold. The scheme of the present application can solve the problem of inaccurate monitoring caused by working condition switching, can capture different types of anomalies, and can reduce the false alarm rate.

Description

Technical Field

[0001] This invention relates to the field of anomaly monitoring technology. More specifically, this invention relates to a method and system for anomaly monitoring in the corrugated cardboard production process. Background Technology

[0002] The corrugated cardboard production process is characterized by multiple steps, continuous operation, and high speed. Its product quality and production efficiency are directly influenced by numerous process parameters, such as temperature, pressure, speed, and glue application rate. These parameters exhibit complex characteristics of high dimensionality, strong coupling, and dynamic time-varying behavior. Therefore, timely detection and diagnosis of abnormal operating conditions are crucial to ensuring stable production. Current technologies for industrial process monitoring typically employ statistical process control (SPC) methods, such as Shewhart control charts and cumulative sum control charts. However, these methods usually assume that the data follows a single Gaussian distribution. This assumption makes them ineffective in handling the multimodal and nonlinear data commonly encountered in corrugated cardboard production, and they are not sensitive enough to small, gradual anomalies, thus failing to meet complex monitoring needs.

[0003] To address these issues, researchers have proposed data-driven monitoring methods, such as Principal Component Analysis (PCA), Support Vector Machines (SVM), and neural networks. These methods train models using a large number of labeled normal and faulty samples, learning process characteristics from massive historical data. This improves the monitoring performance of complex industrial processes to some extent, better handles nonlinear relationships, and offers certain advantages over traditional statistical process control methods.

[0004] However, the aforementioned data-driven monitoring methods still have many shortcomings in practical applications. First, these methods require a large number of labeled normal and faulty samples for model training, but in actual production, faulty samples are usually scarce and difficult to obtain, leading to insufficient model training. Second, these methods are often considered "black box models," lacking physical interpretability in their monitoring results and failing to adapt adequately to dynamic changes in production conditions. Furthermore, production lines need to frequently switch between different production grades based on order demands, leading to multimodal data characteristics. Single-condition models are prone to false alarms or missed alarms due to model mismatch during condition switching. Additionally, using a fixed-length data analysis window makes it difficult to effectively capture instantaneous impacts caused by sudden equipment failures and gradual anomalies caused by component wear. Moreover, monitoring methods based on statistical distance (such as relative entropy KL divergence) fail to distinguish the economic costs corresponding to deviations in different variables, resulting in low correlation between monitoring results and actual production risks. Finally, using fixed thresholds for anomaly detection cannot adapt to changing conditions such as production speed, environmental temperature, and humidity, easily leading to misjudgments. Summary of the Invention

[0005] The purpose of this invention is to propose a method and system for monitoring anomalies in the corrugated cardboard production process, in order to solve the problems of insufficient model training, low monitoring accuracy in multi-brand production mode, inability to effectively capture instantaneous impact anomalies and slow gradual anomalies, and easy misjudgment caused by changes in external conditions in the prior art; to this end, this invention provides solutions in the following two aspects.

[0006] In a first aspect, the present invention provides a method for monitoring anomalies in the corrugated cardboard production process, comprising: Obtain historical normal datasets and real-time data streams of multiple process parameters during corrugated cardboard production; based on the historical normal datasets, for A baseline Gaussian mixture model was constructed for each of the different production grades and operating conditions to obtain... The system calculates a baseline probability density function and a baseline statistical moment vector; it calculates the statistical moment vector of the real-time data stream within a preset time period, and determines the fusion weight of each baseline probability density function based on the statistical moment vector and each baseline statistical moment vector; it calculates the weight entropy based on the fusion weight, and performs weighted fusion on each baseline probability density function to obtain a composite baseline probability density function; it calculates the kurtosis value of the first-order difference sequence of the real-time data stream, and adaptively selects analysis windows of different lengths based on the comparison result of the kurtosis value and a preset kurtosis threshold, and constructs the current probability density function based on the data within the analysis window; it calculates the asymmetric penalty relative entropy between the composite baseline probability density function and the current probability density function; it uses a first judgment threshold when the current equipment operating speed is greater than the speed threshold and the ambient humidity is greater than the humidity threshold, otherwise it uses a second judgment threshold; when the asymmetric penalty relative entropy is greater than the used judgment threshold, it determines that an abnormality has occurred in the production process.

[0007] Preferably, calculating the asymmetric penalty relative entropy between the composite baseline probability density function and the current probability density function includes: In the formula, For asymmetric penalty relative entropy, For the penalty weight function, For the composite baseline probability density function, This is the current probability density function.

[0008] Preferably, the penalty weight function satisfies the expression: In the formula, For the penalty weight function, For data points The physical unit cost of the corresponding process parameters. The average physical unit cost of a historical normal dataset. This is the penalty coefficient.

[0009] Preferably, the above is A baseline Gaussian mixture model was constructed for each of the different production grades and operating conditions, including: using the expectation-maximization method. The algorithm trains a Gaussian mixture model, determines the optimal number of Gaussian components in the Gaussian mixture model using the Bayesian information criterion, and completes the construction of a benchmark Gaussian mixture model under various production grades and operating conditions.

[0010] Preferably, determining the fusion weights of each benchmark probability density function includes: calculating the mean and covariance of the real-time data stream within a preset time period to form a statistical moment vector; and calculating the statistical moment vector and... The Mahalanobis distance between the baseline statistical moment vectors is used; the reciprocal of each Mahalanobis distance is taken and normalized to obtain the fusion weight of each baseline probability density function; the normalization process is as follows: In the formula, For the first The fusion weights of the baseline probability density functions, , These are the statistical moment vectors and the first... , Mahalanobis distance between the baseline statistical moment vectors This represents the total number of baseline probability density functions.

[0011] Preferably, the adaptive selection of analysis windows of different lengths includes: calculating the kurtosis value of the first-order difference sequence of the real-time data stream within the most recent 100 data points; when the kurtosis value is greater than a preset kurtosis threshold, selecting an analysis window of length [missing information]. Analysis window for each data point; when the kurtosis value is not greater than the preset kurtosis threshold, select a length of [length missing]. Analysis window for each data point.

[0012] Preferably, the data points The physical unit cost of the corresponding process parameters is determined as follows: a mapping relationship between process parameters and physical unit costs is established in advance, and the mapping relationship is queried based on the real-time collected process parameter values ​​to determine the data points. Physical unit cost of the corresponding process parameters The physical unit cost is positively correlated with the steam pressure of the paper heater and the glue temperature of the glue applicator.

[0013] Preferably, the penalty coefficient The value is determined based on the comparison between the weighted entropy and the entropy threshold, and is set to the first preset value. Or the second preset value When the weight entropy is greater than the entropy threshold, the second preset value is taken. When the weight entropy is not greater than the entropy threshold, the first preset value is taken. .

[0014] Preferably, the first determination threshold is 2.5; the second determination threshold is 1.8.

[0015] In the second aspect, a corrugated cardboard production process anomaly monitoring system includes: The system includes a processor and a memory, the memory storing computer program instructions for monitoring anomalies in the corrugated board production process, which, when executed by the processor, implement the aforementioned method for monitoring anomalies in the corrugated board production process.

[0016] The beneficial effects of this invention are as follows: 1. This invention constructs benchmark models for different production grades and operating conditions using historical normal datasets, enabling thorough model training. Simultaneously, it weights and fuses real-time data based on their closeness to each benchmark model, generating a composite benchmark that accurately matches the current production state. This solves the monitoring inaccuracy problem caused by model mismatch due to operating condition switching, thereby improving monitoring accuracy under multi-grade production modes. Secondly, this invention adaptively selects appropriate analysis windows by analyzing the statistical characteristics of real-time data streams, enabling rapid response to instantaneous, impactful anomalies and stable detection of slow, gradual anomalies. Furthermore, it employs corresponding judgment criteria based on different current equipment operating speeds and ambient humidity levels to prevent normal fluctuations caused by changes in external conditions from being misjudged as anomalies, thus reducing the false alarm rate and making anomaly detection more reliable.

[0017] 2. In this invention, physical unit cost related to process parameters is introduced as a penalty weight when measuring differences in data distribution. This allows the monitoring results to prioritize anomalies that have a greater impact on production costs, thereby enhancing the economic guidance significance and practical value of the monitoring. Attached Figure Description

[0018] Figure 1 This illustration schematically shows a flowchart of the steps in a method for monitoring anomalies in the corrugated cardboard production process according to this embodiment; Figure 2 The schematic diagram illustrates the structural block diagram of an anomaly monitoring system for corrugated cardboard production process in this embodiment. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] like Figure 1 As shown, the method for monitoring anomalies in the corrugated cardboard production process in this embodiment includes steps S1 to S4: Step S1: Obtain historical normal datasets and real-time data streams of multiple process parameters during corrugated cardboard production; based on the historical normal datasets, for... A baseline Gaussian mixture model was constructed for each of the different production grades and operating conditions to obtain... A baseline probability density function and a baseline statistical moment vector.

[0021] Specifically, data on multiple key process parameters of the corrugated cardboard production line are collected using a Manufacturing Execution System (MES) or a Supervisory Control and Data Acquisition (SCADA) system, forming a multi-dimensional time series. These process parameters include paper strength, glue application rate, steam pressure, glue temperature, and production line speed. Data collected during periods of stable production line operation with consistently high-quality output is used as the historical normal dataset after removing obvious outliers and filling in missing values. Simultaneously, data on the same process parameters is collected from the production line in real-time at a fixed sampling frequency, forming a real-time data stream; in this embodiment, the sampling frequency is once per second. All parameter data in both the historical normal dataset and the real-time data stream undergo Z-score standardization to eliminate the influence of different dimensions between parameter data.

[0022] In one embodiment, the term is A baseline Gaussian mixture model was constructed for each of the different production grades and operating conditions, including: Using Expectation Maximization The algorithm trains a Gaussian mixture model, determines the optimal number of Gaussian components in the Gaussian mixture model using the Bayesian information criterion, and completes the construction of a benchmark Gaussian mixture model under various production grades and operating conditions.

[0023] Specifically, production on the production line , , Taking three different brands of products as examples, It equals 3. (Regarding the brand / grade) We collected 10,000 sets of historical process parameter data generated during continuous operation under stable production conditions. Each set of data included multiple dimensions such as raw paper strength, glue application amount, steam pressure, glue temperature, and production line speed. Using this data, we adopted the expectation-maximization method... Algorithm to train a Gaussian mixture model Multiple Gaussian mixture models were obtained. Using Bayesian information criteria Determine the optimal number of Gaussian components for the Gaussian mixture model; specifically, increase the number of Gaussian components sequentially from 2 to 8, calculate the Bayesian information criterion value for each Gaussian mixture model, and compare the values ​​corresponding to the number of Gaussian components. Value, found The minimum value, the number of Gaussian components corresponding to this minimum value is the grade. The optimal number of Gaussian components in a Gaussian mixture model indicates that the model has achieved the best balance between fitting performance and complexity. Determining the optimal complexity of the Gaussian mixture model is crucial. The baseline Gaussian mixture model is composed of the optimal number of Gaussian components, which is then used to determine the grade. The baseline Gaussian mixture model. For grade , The above steps are then repeated to establish the optimal benchmark Gaussian mixture model for the three different production grades and operating conditions.

[0024] After training After establishing the baseline Gaussian mixture models for different production grades and operating conditions, the baseline probability density functions for the corresponding grade and operating conditions are obtained. The baseline probability density function is a weighted sum of multiple Gaussian distributions. Simultaneously, the statistical moments of all historical normal data for each brand are calculated, forming a statistical moment vector. The statistical moment vector contains two elements, which are the mean and covariance of all historical normal data for the corresponding brand.

[0025] Step S2: Calculate the statistical moment vector of the real-time data stream within a preset time period; determine the fusion weight of each benchmark probability density function based on the statistical moment vector and each benchmark statistical moment vector; calculate the weight entropy according to the fusion weight, and perform weighted fusion on each benchmark probability density function to obtain a composite benchmark probability density function.

[0026] In one embodiment, determining the fusion weights of each baseline probability density function includes: Within a preset time period, the mean and covariance of the real-time data stream are calculated to form a statistical moment vector; the statistical moment vector and... Mahalanobis distance between the reference statistical moment vectors; Take the reciprocal of each Mahalanobis distance and normalize it to obtain the fusion weight of each baseline probability density function; The normalization process is as follows: In the formula, For the first The fusion weights of the baseline probability density functions, , These are the statistical moment vectors and the first... , Mahalanobis distance between the baseline statistical moment vectors This represents the total number of baseline probability density functions.

[0027] Specifically, a 60-second sliding time window is set as the preset time period on the real-time data stream. For the data within each window, the mean and covariance are calculated to form the statistical moment vector at the current moment. Understandably, multiple process parameter data are continuously collected during real-time production, with a 60-second cycle. At the end of the current 60-second cycle, the mean and covariance of all real-time data within that period are immediately calculated and combined to form a real-time statistical moment vector. The statistical moment vector is then calculated separately. and The Mahalanobis distance between the reference statistical moment vectors is obtained. The reciprocal of each Mahalanobis distance is used, and the reciprocal is normalized. for The sum of the reciprocals of the Mahalanobis distances, after normalization, represents the fusion weights corresponding to the baseline probability density function. The sum of all fusion weights is 1. For example, still using production line production... , , Taking three different product brands as examples, when designing the brand... , , After pre-constructing a benchmark Gaussian mixture model, the grade is calculated. , , The statistical moment vectors are respectively , , The three distance values ​​obtained are as follows: , , The calculated distance value , , Taking 2, 8, and 5 as examples, the reciprocals of each distance are 0.5, 0.125, and 0.2, respectively, with a sum of 0.825. These are then normalized to obtain the fusion weights. Approximately 0.61 Approximately 0.15 It is approximately 0.24.

[0028] In one embodiment, the weight entropy satisfies the expression: In the formula, For weighted entropy, For the first The fusion weights of the baseline probability density functions, It is a logarithmic function with base 2. This represents the total number of baseline probability density functions. Weight entropy is an indicator of the uniformity of the distribution of this set of fused weights; a higher entropy value indicates greater uncertainty.

[0029] In one embodiment, the composite baseline probability density function satisfies the expression: In the formula, For the composite baseline probability density function, For the first The fusion weights of the baseline probability density functions, For the first A baseline probability density function, This represents the total number of baseline probability density functions.

[0030] Step S3: Calculate the kurtosis value of the first-order difference sequence of the real-time data stream; based on the comparison result of the kurtosis value and the preset kurtosis threshold, adaptively select analysis windows of different lengths; construct the current probability density function based on the data within the analysis window; calculate the asymmetric penalty relative entropy between the composite benchmark probability density function and the current probability density function.

[0031] In one embodiment, the adaptive selection of analysis windows of different lengths includes: Calculate the kurtosis of the first-order difference sequence of the real-time data stream over the most recent 100 data points; When the kurtosis value is greater than the preset kurtosis threshold, select a length of... Analysis window for each data point; When the kurtosis value is not greater than the preset kurtosis threshold, select a length of... Analysis window for each data point.

[0032] Specifically, the real-time data stream is continuously monitored. At any given time, the latest 100 consecutive data points are taken. First-order differencing is performed on the data, that is, the data points from the previous time step are subtracted from the current data points, resulting in a difference sequence containing 99 values. This difference sequence effectively amplifies abrupt changes in the data. The kurtosis of the difference sequence is calculated. A kurtosis threshold is preset, empirically set to 6. Based on the comparison between the calculated kurtosis value and the preset threshold, analysis windows of different lengths are selected, including short analysis windows. Length Analysis Window Among them, short analysis window Smaller than long analysis window When the calculated kurtosis value exceeds the preset kurtosis threshold, it indicates that the data has experienced drastic fluctuations or shocks, suggesting that the production process may be unstable. In this case, a short analysis window with a length of 50 data points is automatically selected. Subsequent anomaly detection is performed to quickly respond to such changes and capture the anomaly. Conversely, when the calculated kurtosis value is not greater than the preset kurtosis threshold, it indicates that the data stream is relatively stable or that there is a slow-changing anomaly. In this case, a long analysis window with a length of 200 data points is automatically selected. This allows for the accumulation of sufficient data to detect subtle changes and obtain smoother, more statistically representative analytical results.

[0033] In one embodiment, kernel density estimation is used. The method involves performing nonparametric probability density estimation on the data within the selected analysis window to construct the probability density function of the current working condition, i.e., the current probability density function.

[0034] In one embodiment, calculating the asymmetric penalty relative entropy between the composite baseline probability density function and the current probability density function includes: ; In the formula, For asymmetric penalty relative entropy, For the penalty weight function, For the composite baseline probability density function, This is the current probability density function.

[0035] The penalty weight function satisfies the expression: ; In the formula, For the penalty weight function, For data points The physical unit cost of the corresponding process parameters. The average physical unit cost of a historical normal dataset. This is the penalty coefficient.

[0036] Specifically, integration is achieved by summing the data sampling points in actual calculations. For a multidimensional data point... Physical unit cost The unit product energy consumption and raw material cost are calculated using a pre-established cost model, representing the values ​​of the obtained process parameters. It is the average output value of the cost model across all historical normal datasets. The term is a core component of relative entropy (also often called KL divergence). It is a logarithmic function with the natural constant e as the base, which calculates two points distributed across the data points. Information discrepancies.

[0037] In one embodiment, the data points The physical unit cost of the corresponding process parameters is determined in the following way: A mapping relationship between process parameters and physical unit cost is pre-established. The mapping relationship is then queried based on the real-time collected process parameter values ​​to determine the data points. Physical unit cost of the corresponding process parameters ; The physical unit cost is positively correlated with the steam pressure of the paper heater and the glue temperature of the glue applicator.

[0038] Specifically, before implementation, a cost model needs to be established based on production experience and material consumption data. This model clearly indicates how key process parameters affect the production cost per unit product. The cost function, i.e., the pre-defined mapping relationship, is then derived through analysis. The cost function shows that the higher the steam pressure and adhesive temperature, the greater the energy and raw material consumption, and the higher the unit cost.

[0039] During the production process, each data point is collected in real time. The various process parameters. For example, at a certain moment, the data points collected... The included paper heater has a steam pressure of 1.5 MPa, and the glue temperature of the gluing machine is 70 degrees Celsius. Data points are calculated using a preset mapping relationship. Physical unit cost of the corresponding process parameters ,get The value of 35.3 represents the instantaneous production cost under the given process parameters.

[0040] In one embodiment, the penalty coefficient The value is determined based on the comparison between the weighted entropy and the entropy threshold, and is set to the first preset value. Or the second preset value When the weight entropy is greater than the entropy threshold, the second preset value is taken. When the weight entropy is not greater than the entropy threshold, the first preset value is taken. .

[0041] Specifically, based on experience, an entropy threshold of 0.8 is preset, which is the first preset value. The value is 0.5, the second preset value. The value is 1.5. The fusion weight is determined based on the different production grades and operating conditions. Calculate the weight entropy Then, based on the weight entropy The penalty coefficient is taken from the comparison result with the entropy threshold. The value is the first preset value. Or the second preset value ,and When the calculated weighted entropy is greater than the entropy threshold, it indicates that it is impossible to clearly determine which production grade the current condition belongs to, and the uncertainty of the working condition identification is very high. In this case, the second preset value is selected. , This serves as a larger penalty coefficient to enhance sensitivity to deviations from the target state; when the calculated weighted entropy is not greater than the entropy threshold, it indicates that the operating condition is clearly identified, and the first preset value is selected at this time. , , as a smaller penalty coefficient.

[0042] Step S4: When the current equipment operating speed is greater than the speed threshold and the ambient humidity is greater than the humidity threshold, the first judgment threshold is used; otherwise, the second judgment threshold is used. When the relative entropy of the asymmetric penalty is greater than the judgment threshold used, it is determined that an abnormality has occurred in the production process.

[0043] Specifically, two key environmental and equipment status thresholds are pre-set. In this embodiment, the equipment operating speed and workshop ambient humidity are selected based on experience, with the preset speed threshold fixed at 200 meters per minute and the humidity threshold fixed at 85%RH. Simultaneously, two sets of corresponding anomaly detection thresholds are preset based on experience, with the first threshold set at 2.5 and the second at 1.8. During equipment operation, the equipment operating speed and workshop ambient humidity are continuously acquired from sensors on the production line. It is determined whether the current operating speed exceeds the speed threshold and whether the ambient humidity exceeds the humidity threshold, thus deciding which threshold to use. Finally, the comparison between the asymmetric penalty relative entropy and the adopted threshold determines whether an anomaly has occurred in the production process. For example, if the monitored equipment operating speed is 210 meters per minute and the current workshop humidity is 90%, it is determined that the current equipment operating speed is greater than the speed threshold and the current workshop humidity is greater than the humidity threshold. Both conditions are met simultaneously, and this high-speed, high-humidity condition is prone to large fluctuations in process parameters, considered a high-risk production situation. Therefore, a stricter first judgment threshold is used for anomaly detection. The calculated asymmetric penalty relative entropy is compared with the first judgment threshold; if it is greater, the production process is determined to be abnormal. Conversely, under other conditions, if the monitored equipment operating speed is less than the speed threshold or the current workshop humidity is less than the humidity threshold, neither condition is met, and the production process should be more stable. Therefore, a more lenient second judgment threshold is used for anomaly detection. The calculated asymmetric penalty relative entropy is compared with the second judgment threshold; if it is greater, the production process is determined to be abnormal.

[0044] This invention also provides an anomaly monitoring system for the corrugated cardboard production process. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions for monitoring anomalies in the corrugated cardboard production process. When the computer program instructions are executed by the processor, the above-described method for monitoring anomalies in the corrugated cardboard production process according to the present invention is implemented.

[0045] 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 therefore will not be described in detail here.

[0046] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined 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 desired 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. Any application or module described in this invention can be implemented by computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0047] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0048] While various embodiments of the invention have been shown and described in this specification, 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 occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for monitoring anomalies in the corrugated cardboard production process, characterized in that, include: Acquire historical normal datasets and real-time data streams of multiple process parameters during corrugated cardboard production; Based on the aforementioned historical normal dataset, for A baseline Gaussian mixture model was constructed for each of the different production grades and operating conditions to obtain... One baseline probability density function and one baseline statistical moment vector; Calculate the statistical moment vector of the real-time data stream within a preset time period. Based on the statistical moment vector and each benchmark statistical moment vector, determine the fusion weight of each benchmark probability density function. Calculate the weight entropy according to the fusion weight, and perform weighted fusion of each benchmark probability density function to obtain a composite benchmark probability density function. Calculate the kurtosis value of the first-order difference sequence of the real-time data stream, and adaptively select analysis windows of different lengths based on the comparison result of the kurtosis value and the preset kurtosis threshold, and construct the current probability density function based on the data in the analysis window; Calculating the asymmetric penalty relative entropy between the composite baseline probability density function and the current probability density function includes: ； In the formula, For asymmetric penalty relative entropy, For the penalty weight function, For the composite baseline probability density function, The current probability density function; The penalty weight function satisfies the expression: ； In the formula, For the penalty weight function, For data points The physical unit cost of the corresponding process parameters. The average physical unit cost of a historical normal dataset. This is the penalty coefficient; Penalty coefficient The value is determined based on the comparison between the weighted entropy and the entropy threshold, and is set to the first preset value. Or the second preset value When the weight entropy is greater than the entropy threshold, the second preset value is taken. When the weight entropy is not greater than the entropy threshold, the first preset value is taken. ; When the current equipment operating speed is greater than the speed threshold and the ambient humidity is greater than the humidity threshold, the first judgment threshold is used; otherwise, the second judgment threshold is used. When the relative entropy of the asymmetric penalty is greater than the judgment threshold used, the production process is determined to be abnormal.

2. The method for monitoring anomalies in the corrugated cardboard production process according to claim 1, characterized in that, The above is A baseline Gaussian mixture model was constructed for each of the different production grades and operating conditions, including: Using Expectation Maximum The algorithm trains a Gaussian mixture model, and determines the optimal number of Gaussian components in the Gaussian mixture model through the Bayesian information criterion, thus completing the construction of the benchmark Gaussian mixture model under the working conditions of each production grade.

3. The method for monitoring anomalies in the corrugated cardboard production process according to claim 1, characterized in that, The determination of the fusion weights for each baseline probability density function includes: Within a preset time period, the mean and covariance of the real-time data stream are calculated to form a statistical moment vector; the statistical moment vector and... Mahalanobis distance between the reference statistical moment vectors; Take the reciprocal of each Mahalanobis distance and normalize it to obtain the fusion weight of each baseline probability density function; The normalization process is as follows: In the formula, For the first The fusion weights of the baseline probability density functions, , These are the statistical moment vectors and the first... , Mahalanobis distance between the baseline statistical moment vectors This represents the total number of baseline probability density functions.

4. The method for monitoring anomalies in the corrugated cardboard production process according to claim 1, characterized in that, The adaptive selection of analysis windows of different lengths includes: Calculate the kurtosis of the first-order difference sequence of the real-time data stream over the most recent 100 data points; When the kurtosis value is greater than the preset kurtosis threshold, select a length of... Analysis window for each data point; When the kurtosis value is not greater than the preset kurtosis threshold, select a length of... Analysis window for each data point.

5. The method for monitoring anomalies in the corrugated cardboard production process according to claim 1, characterized in that, The data points The physical unit cost of the corresponding process parameters is determined in the following way: A mapping relationship between process parameters and physical unit cost is pre-established. The mapping relationship is then queried based on the real-time collected process parameter values ​​to determine the data points. Physical unit cost of the corresponding process parameters ; The physical unit cost is positively correlated with the steam pressure of the paper heater and the glue temperature of the glue applicator.

6. The method for monitoring anomalies in the corrugated cardboard production process according to claim 1, characterized in that, The first determination threshold is 2.5; the second determination threshold is 1.

8.

7. An anomaly monitoring system for corrugated cardboard production process, characterized in that, include: The processor and memory, wherein the memory stores computer program instructions for monitoring anomalies in the corrugated board production process, and when the computer program instructions are executed by the processor, implement a method for monitoring anomalies in the corrugated board production process according to any one of claims 1-6.

Images