Processing and filling monitoring method based on multi-source sensing data

Abstract

The invention relates to the technical field of data processing, in particular to a processing and filling monitoring method based on multi-source sensing data, and the method comprises the steps: carrying out the weighted fusion of all sensor data in a processing and filling process according to the correlation between the sensor data and the filling quality, and obtaining a real-time sequence of fusion data; calculating the local deviation of the real-time sequence and the standard sequence in each time period, and obtaining the noise probability of each time period in the real-time sequence; and performing weighted summation on the local deviation of each time period according to the noise probability to obtain a global deviation, and responding to the situation that the global deviation is greater than a global threshold, determining that the monitoring result is abnormal, otherwise, determining that the monitoring result is normal. According to the technical scheme, the influence of noise on the monitoring result can be eliminated, and the monitoring result of processing and filling can be accurately obtained.

Description

Technical Field

[0001] This application relates to the field of data processing technologies, and particularly to a processing and filling monitoring method based on multi-source sensing data. Background Art

[0002] The processing and filling process is a key link in the production process of chemical reagents, which is directly related to the production quality during the product packaging process. To ensure the production quality of chemical reagents, how to accurately monitor the key parameters in processing and filling has become an urgent problem to be solved.

[0003] Currently, the patent application document with the publication number CN116933210A discloses a monitoring method and system for food processing and filling equipment based on multi-dimensional sensors. The method includes: obtaining the monitoring data of the food processing and filling equipment, and constructing a filling area monitoring matrix using the monitoring data; obtaining the internal deviation coefficient of the filling area according to the filling area monitoring matrix; obtaining the appearance matching coefficient of the filling area according to the filling area monitoring matrix; obtaining the comprehensive quality deviation index according to the internal deviation coefficient and the appearance matching coefficient of the filling area; obtaining the comprehensive deviation similarity according to the comprehensive quality deviation index; obtaining the filling abnormal risk index according to the comprehensive quality deviation index; obtaining the filling area abnormal confidence coefficient according to the filling abnormal risk index and the comprehensive deviation similarity; and completing the monitoring of the filling equipment according to the filling area abnormal confidence coefficient.

[0004] The above method obtains the internal deviation coefficient and the appearance matching coefficient based on the monitoring data, and realizes the monitoring of the filling equipment by comprehensively considering the internal deviation coefficient and the appearance matching coefficient. Among them, the internal deviation coefficient is obtained according to the weight data and the liquid level height data, and the appearance matching coefficient is obtained according to the chromaticity data. However, the working environment of the processing and filling equipment is complex, and during the process of obtaining the monitoring data, noise data irrelevant to the monitoring process will be introduced, resulting in inaccurate monitoring results of processing and filling. Summary of the Invention

[0005] To solve the technical problem of inaccurate monitoring results of processing and filling, this application provides a processing and filling monitoring method based on multi-source sensing data, which can eliminate the influence of noise on the monitoring results and accurately obtain the monitoring results of processing and filling.

[0006] In the first aspect of this application, a processing and filling monitoring method based on multi-source sensing data is provided. The monitoring method includes: weighted-fusing each sensor data during the processing and filling process according to the correlation between the sensor data and the filling quality to obtain a real-time sequence of the fused data; calculating the local deviation between the real-time sequence and the standard sequence in each time period to obtain the noise probability in each time period of the real-time sequence, including: obtaining the sample sequence of the fused data and the filling quality during the historical canning process, and equally spacing the sample sequence into multiple time periods; calculating the local deviation between the sample sequence and the standard sequence in each time period, and taking the time period with the local deviation greater than the local threshold as the difference segment. In response to the number of difference segments being 1, taking the timestamp of the difference segment and the difference sequence between the sample sequence and the standard sequence within the difference segment as a set of training data; in response to the absolute value of the difference between the filling quality and the standard quality being greater than the difference threshold, the noise label of the training data is 0, otherwise, the noise label of the training data is 1; inputting the training data into a binary classification model to output the noise probability, and training the binary classification model based on the cross-entropy function between the noise probability and the noise label; inputting the timestamp and the difference sequence in each time period of the real-time sequence into the trained binary classification model to output the noise probability in each time period of the real-time sequence; weighted-summing the local deviations in each time period according to the noise probability to obtain the global deviation. In response to the global deviation being greater than the global threshold, the monitoring result is abnormal, otherwise the monitoring result is normal.

[0007] During the processing and filling process, a variety of sensor data can be collected at each moment. According to the correlation between each sensor data and the filling quality, the sensor data at each moment is weighted to obtain the fused data at each moment of the processing and filling process, and then a real-time sequence of the fused data is obtained; the real-time sequence of the fused data is equally spaced into multiple time periods, and the local deviation between the real-time sequence and the standard sequence in each time period is calculated, as well as the noise probability in each time period of the real-time sequence. The noise probability can reflect the degree of influence of the fluctuation of the real-time sequence in the corresponding time period on the filling quality. The greater the noise probability, the smaller the influence of the fluctuation of the real-time sequence in the corresponding time period on the filling quality, and the smaller the noise probability, the greater the influence of the fluctuation of the real-time sequence in the corresponding time period on the filling quality; weighted-summing the local deviations in each time period according to the noise probability to obtain the global deviation. The global deviation eliminates the influence of noise data irrelevant to the canned quality monitoring on the monitoring result, and by comparing the global deviation and the global threshold, the monitoring result of the processing and filling can be accurately obtained.

[0008] Preferably, the sensor data and the calculation method of the correlation between the filling quality includes: the sensor data Take the DTW distance between the historical sequences of each sensor data other than [sensor data] as the clustering distance, and perform clustering on each sample data to obtain multiple clustering clusters. The sample data includes the historical sequences of each sensor data and the canning quality during the historical canning process; within any clustering cluster, calculate the average sequence of the sensor data of each historical sequence, and take the variance of the DTW distance between each historical sequence of the sensor data and the average sequence as the first variance, and calculate the absolute value of the difference between the first variance and the canning quality variance in the clustering cluster; weight and sum the absolute values of the differences of each clustering cluster according to the proportion of the sample data in the clustering cluster. The correlation between the sensor data and the filling quality is negatively correlated with the weighted sum result.

[0009] Cluster the sample data according to the DTW distance between the historical sequences of each sensor data other than [sensor data] to obtain multiple clustering clusters. The sensor data in the clustering cluster other than [sensor data] has basically the same historical sequences; calculate the correlation between the sensor data other than [sensor data] and the filling quality in the clustering cluster, and exclude the influence of each sensor data other than [sensor data] to ensure the accuracy of the correlation. to ensure the accuracy of the correlation.

[0010] Preferably, the correlation between the sensor data and the filling quality is: , is any minimum value, is the weighted sum result of the sensor data .

[0011] Preferably, the weighted sum result of the sensor data satisfies the relational expression: ; is the total number of clustering clusters, is the number of sample data in the clustering cluster , is the number of all sample data, and are respectively the first variance and the canning quality variance of the clustering cluster .

[0012] Each clustering cluster can obtain an absolute value of the difference between the first variance and the canning quality variance. The larger the sample data in the clustering cluster, the more the obtained absolute value of the difference can truly reflect the sensor data And the correlation with the filling quality. Therefore, the absolute value of the difference of each clustering cluster is weighted by the number of sample data within the clustering cluster to ensure the accuracy of the correlation.

[0013] Preferably, at the moment The fused data Is: ; Is the type of sensor data, Is the sensor data Of the correlation, Is the sum of the correlations of all sensor data, Is at the moment Sensor data Of the dimensionless value.

[0014] Preferably, before obtaining the real-time sequence of the fused data, the monitoring method further includes: performing denoising processing on the sensor data by using Gaussian filtering.

[0015] Preferably, the method for obtaining the standard sequence includes: taking the sample sequence with the filling quality greater than the quality threshold as the candidate sequence, calculating the average value at any moment in all candidate sequences to obtain the standard sequence; the standard quality is the average filling quality of each candidate sequence.

[0016] Preferably, the global deviation Satisfies the relational expression: , Is the number of time periods, Is the time period in the real-time sequence Of the noise probability, Is the time period Of the local deviation.

[0017] The global deviation eliminates the influence of noise data irrelevant to the canned quality monitoring on the monitoring result, so that the global deviation can accurately reflect the difference between the real-time sequence of the fused data and the standard sequence. The greater this difference, the greater the possibility of an abnormality in the current processing and filling process.

[0018] Preferably, training the binary classification model based on the cross-entropy function between the noise probability and the noise label includes: calculating the proportion of the number of training data in each time period in the total amount of training data, and taking the difference between 1 and the proportion as the normalized weight of the corresponding time period; calculating the cross-entropy function between the noise probability and the noise label, and training the binary classification model based on the product of the cross-entropy function and the normalized weight of the time period in the training data.

[0019] To ensure that the binary classification model can accurately obtain the noise probability at any time period, it is necessary to pay attention to whether the quantity of training data in different time periods is balanced during the training process, allocate a larger normalization weight to the training data corresponding to the time period with a smaller proportion, alleviate the problem of unbalanced quantity of training data in each time period, and enable the binary classification model to accurately obtain the noise probability at any time period.

[0020] Preferably, the binary classification model is a fully connected neural network or a logical classification model.

[0021] The technical solution of this application has the following beneficial technical effects: During the processing and filling process, various sensor data can be collected at each moment. According to the correlation between each sensor data and the filling quality, the sensor data at each moment is weighted to obtain the fusion data at each moment of the processing and filling process, and then the real-time sequence of the fusion data is obtained. The real-time sequence of the fusion data is equally spaced into multiple time periods, and the local deviation between the real-time sequence and the standard sequence within each time period, as well as the noise probability of each time period in the real-time sequence, are calculated. The noise probability can reflect the degree of influence of the fluctuation of the real-time sequence in the corresponding time period on the filling quality. The greater the noise probability, the smaller the influence of the fluctuation of the real-time sequence in the corresponding time period on the filling quality, and the smaller the noise probability, the greater the influence of the fluctuation of the real-time sequence in the corresponding time period on the filling quality. The local deviations of each time period are weighted and summed according to the noise probability to obtain the global deviation. The global deviation eliminates the influence of noise data irrelevant to the canned quality monitoring on the monitoring result. By comparing the global deviation and the global threshold, the monitoring result of the processing and filling can be accurately obtained. Description of the Drawings

[0022] Figure 1 is a flowchart of a processing and filling monitoring method based on multi-source sensing data according to an embodiment of the present application. Detailed Embodiments

[0023] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0024] According to the first aspect of the present application, the present application provides a processing and filling monitoring method based on multi-source sensing data. Figure 1 is a flowchart of a processing and filling monitoring method based on multi-source sensing data according to an embodiment of the present application. As Figure 1 shown, the processing and filling monitoring method based on multi-source sensing data includes steps S101 to S103, which are described in detail below.

[0025] S101, weighted fusion of sensor data during the processing and filling process is performed according to the correlation between the sensor data and the filling quality to obtain a real-time sequence of the fused data.

[0026] In one embodiment, after the end of a historical canning process, a historical sequence of sensor data during the historical canning process is collected, and the canning quality of the historical canning process is manually marked. The canning quality adopts a scoring system, and the value range of the canning quality is from 0 to 10. The larger the value, the better the canning quality.

[0027] Among them, the sensor data includes canning speed, canning pressure, canning weight, etc.

[0028] In this way, the historical sequence of each sensor data corresponds to a canning quality, and the canning quality is a score from 0 to 10. Taking the historical sequence of sensor data and the canning quality during a historical canning process as sample data, the correlation between the sensor data and the canning quality can be calculated based on the sample data.

[0029] Specifically, the calculation method of the correlation between the sensor data and the canning quality includes: taking the DTW distance between the historical sequences of sensor data other than as the clustering distance, performing clustering processing on each sample data to obtain multiple clustering clusters; within any clustering cluster, calculating the average sequence of the historical sequences of the sensor data , taking the variance of the DTW distance between the historical sequences of the sensor data and the average sequence as the first variance, and calculating the absolute value of the difference between the first variance and the canning quality variance in the clustering cluster; performing weighted summation on the absolute values of the differences of each clustering cluster according to the proportion of the sample data in the clustering cluster, and the correlation between the sensor data and the canning quality is negatively correlated with the weighted summation result.

[0030] Among them, the Kmeans clustering algorithm can be used to perform clustering processing on the sample data to obtain multiple clustering clusters. A clustering cluster includes at least one sample data; it can be understood that among the multiple sample data in a clustering cluster, the historical sequences of sensor data other than are basically the same. The first variance is used to characterize the degree of difference of the historical sequences of the sensor data in the clustering cluster. The larger the first variance, the greater the degree of difference between the historical sequences of the sensor data in the clustering cluster; the canning quality variance is used to characterize the degree of difference of the canning quality in the clustering cluster. When the first variance is small and the canning quality variance is large, it means that the difference in the canning quality is not caused by the difference in the historical sequences of the sensor data , and the sensor data has a relatively small correlation with the filling quality; when the first variance is large and the canning quality variance is small, it indicates that the sensor data the differences in the historical sequence do not cause differences in the filling quality, and the sensor data also has a relatively small correlation with the filling quality; when the first variance is basically the same as the canning quality variance, it indicates that the sensor data the differences in the historical sequence can cause differences in the filling quality, and the sensor data has a relatively large correlation with the filling quality. Therefore, the correlation is negatively correlated with the absolute value of the difference between the first variance and the canning quality variance.

[0031] Specifically, the correlation between the sensor data and the filling quality is: , is any minimum value, is the weighted summation result of the sensor data . Among them, the minimum value is used to avoid the situation where the denominator is 0. In the embodiments of the present application, the value of the minimum value is 0.1.

[0032] Among them, the weighted summation result of the sensor data satisfies the relational expression: ; is the total number of clustering clusters, is the number of sample data in the clustering cluster , is the number of all sample data, and are respectively the first variance and the canning quality variance of the clustering cluster .

[0033] In this way, in the process of calculating the correlation between the sensor data and the filling quality, the sample data is clustered according to the DTW distance between the historical sequences of the sensor data other than the sensor data to obtain multiple clustering clusters, and the historical sequences of the sensor data other than the sensor data in the clustering cluster are basically the same; the correlation between the sensor data and the filling quality is calculated in the clustering cluster, excluding the influence of the sensor data other than the sensor data to ensure the accuracy of the correlation.

[0034] In one embodiment, after obtaining the correlation between each sensor data and the filling quality, the sensor data at any moment during the processing and filling process is weighted and fused according to the correlation to obtain the fused data at any moment, and then a real-time sequence of the fused data during the processing and filling process is obtained.

[0035] Specifically, the fused data at time is: ; is the type of sensor data, is the correlation of the sensor data , is the sum of the correlations of all sensor data, is the dimensionless value of the sensor data at time . Among them, the dimension of the sensor data can be eliminated by normalization operation or standardization operation to obtain the dimensionless value.

[0036] It should be noted that since the fused data is the weighted sum of the sensor data, and environmental noise is inevitably introduced during the acquisition of the sensor data, therefore, to ensure that the fused data can accurately represent the filling quality, before obtaining the real-time sequence of the fused data, the monitoring method further includes: denoising the sensor data using Gaussian filtering.

[0037] In this way, the sensor data at any moment during the processing and filling process is weighted and fused according to the correlation to obtain the fused data at any moment. The fused data can retain the sensor data that effectively represents the filling quality, and a real-time sequence of the fused data is obtained.

[0038] S102, calculate the local deviation between the real-time sequence and the standard sequence in each time period, and obtain the noise probability in each time period of the real-time sequence.

[0039] In one embodiment, since the real-time sequence is the result of weighted fusion of sensor data, the change of the sensor data will cause fluctuations in the real-time sequence. However, not all fluctuations in the real-time sequence will affect the filling quality. Taking the filling speed during the processing and canning of chemical reagents as an example, in the early stage of canning, since there is less chemical reagent in the can at this time, even if there are large fluctuations in the filling speed, it will not cause a significant decrease in the canning quality. Therefore, the fluctuations in the filling speed at this time do not affect the filling quality, that is to say, relative to the filling quality monitoring, the fluctuations in the filling speed at this time are noise data; while in the later stage of canning, since the chemical reagent in the can basically reaches the canning standard at this time, even if there are small fluctuations in the filling speed, it will cause a significant decrease in the canning quality. Therefore, the fluctuations in the filling speed at this time have a significant impact on the canning quality, that is to say, relative to the filling quality monitoring, the fluctuations in the filling speed at this time are non-noise data.

[0040] To avoid the influence of noise data in the real-time sequence on the canned quality detection, it is necessary to detect the real-time sequence to obtain the noise level of each time period in the implementation sequence.

[0041] Specifically, obtaining the noise probability of each time period in the real-time sequence includes: obtaining the sample sequence of the fused data and the filling quality during the historical canning process, and equally spacing the sample sequence into multiple time periods; calculating the local deviation of each time period between the sample sequence and the standard sequence, and taking the time period with the local deviation greater than the local threshold as the difference segment. In response to the number of difference segments being 1, taking the timestamp of the difference segment and the difference sequence between the sample sequence and the standard sequence within the difference segment as a set of training data; in response to the absolute value of the difference between the filling quality and the standard quality being greater than the difference threshold, the noise label of the training data is 0, otherwise, the noise label of the training data is 1; inputting the training data into a binary classification model to output the noise probability, and training the binary classification model based on the cross-entropy function between the noise probability and the noise label; inputting the timestamp and the difference sequence of each time period in the real-time sequence into the trained binary classification model to output the noise probability of each time period in the real-time sequence.

[0042] Among them, the sample data is the historical sequence of each sensor data and the canned quality during a historical canning process. The historical sequences of each sensor data are weighted and fused according to the correlation between the sensor data and the canned quality to obtain the sample sequence of the fused data, that is, one sample sequence of the fused data can be collected during one historical canning process. Since the time length of one canning process is the same, the lengths of each sample sequence and the real-time sequence are the same, and the number of time periods they contain is also the same.

[0043] Among them, the local deviation is the Euclidean distance or the DTW distance between the sample sequence and the standard sequence in the corresponding time period; the values of the local threshold and the difference threshold are both 1; compared with the standard sequence, only the difference segment has differences in the sample sequence. When the absolute value of the difference between the filling quality and the standard quality is greater than the difference threshold, it means that the difference segment can cause changes in the filling quality. Therefore, the difference segment is non-noise data for filling quality monitoring. Thus, when the absolute value of the difference between the filling quality and the standard quality is greater than the difference threshold, the noise label of the training data is 0; similarly, when the absolute value of the difference between the filling quality and the standard quality is not greater than the difference threshold, it means that the difference segment does not cause changes in the filling quality. Therefore, the difference segment is noise data for filling quality monitoring, and the noise label of the training data is 1.

[0044] Among them, the binary classification model is used to map the time stamps and difference sequences of any time period into classification results of whether it is noise. The binary classification model can be an existing binary classification model such as a fully connected neural network or a logistic classification model. To ensure that the binary classification model can accurately obtain the noise probability of any time period, it is necessary to pay attention to whether the number of training data in different time periods is balanced during the training process. For example, if all the difference segments in the training data are in the time period , then the trained binary classification model can accurately obtain the noise probability of the time period , while for other time periods outside the time period , the accuracy of the noise probability cannot be guaranteed. Therefore, the difference in the number of training data in different time periods needs to be considered during the training of the binary classification model. Training the binary classification model based on the cross-entropy function between the noise probability and the noise label includes: calculating the proportion of the number of training data in each time period in the total amount of training data, and taking the difference between 1 and the proportion as the normalized weight of the corresponding time period; calculating the cross-entropy function between the noise probability and the noise label, and training the binary classification model based on the product of the cross-entropy function and the normalized weight of the time period in the training data.

[0045] In this way, during the training of the binary classification model, a larger normalized weight is assigned to the training data corresponding to the time period with a smaller proportion, alleviating the problem of unbalanced number of training data in each time period, so that the binary classification model can accurately obtain the noise probability of any time period.

[0046] In other embodiments, sample data can also be obtained through an experimental method to obtain the training data of each time period. For example, to obtain the training data of the time period , adjust the sensor data of the time period in the standard sequence, so that the sample sequence fluctuates in the time period , that is, the time period is the difference segment, so as to obtain the training data of the time period , thereby alleviating the problem of unbalanced number of training data in each time period.

[0047] It should be noted that the standard sequence is the time sequence of the fusion data corresponding to when the canned quality meets the quality requirements; it can be selected by those skilled in the art from the historical canning process, and the standard quality is the canned quality corresponding to the standard sequence. In other embodiments, the method for obtaining the standard sequence includes: taking the sample sequence with the canned quality greater than the quality threshold as the candidate sequence, and calculating the average value at any moment in all candidate sequences to obtain the standard sequence; the standard quality is the average canned quality of each candidate sequence.

[0048] In this way, according to the timestamps of each time period and the difference sequence between the real-time sequence and the standard sequence within the time period, the noise probability of each time period in each real-time sequence is obtained, realizing the accurate measurement of the noise probability of each time period. The larger the noise probability, the smaller the impact of the corresponding time period's fluctuation on the filling quality. When obtaining the monitoring result, the local deviation of the corresponding time period should be ignored to improve the accuracy of the monitoring result.

[0049] S103, weighted sum the local deviations of each time period according to the noise probability to obtain the global deviation. In response to the global deviation being greater than the global threshold, the monitoring result is abnormal; otherwise, the monitoring result is normal.

[0050] In one embodiment, the global deviation satisfies the relational expression: , is the number of time periods, is the noise probability of the time period in the real-time sequence, is the time period 's local deviation.

[0051] In this way, the global deviation eliminates the influence of noise data irrelevant to the canned quality monitoring on the monitoring result, enabling the global deviation to accurately reflect the difference between the real-time sequence and the standard sequence of the fused data. The greater this difference, the greater the possibility of an abnormality in the current processing and filling process.

[0052] Comparing the global deviation greater than the global threshold, in response to the global deviation being greater than the global threshold, it indicates that the degree of deviation of the real-time sequence of the fused data from the standard sequence in this processing and filling process is greater, and the monitoring result is abnormal; otherwise, the monitoring result is normal. Among them, the value of the global threshold is 2.

[0053] In this way, the monitoring of the canned quality in the processing and filling process is realized. When the monitoring result is abnormal, it means that the canned equipment needs to be repaired to ensure that a better filling quality can be obtained in the next processing and filling process.

[0054] It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several deformations and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of this application patent should be subject to the appended claims.

Claims

1. A processing and filling monitoring method based on multi-source sensing data, characterized in that, The monitoring method includes: weighted fusion of sensor data during the processing and filling process according to the correlation between the sensor data and the filling quality to obtain a real-time sequence of the fused data; Calculating the local deviation between the real-time sequence and the standard sequence in each time period to obtain the noise probability of each time period in the real-time sequence, including: obtaining the sample sequence of the fused data and the filling quality during the historical canning process, and equally spacing the sample sequence into multiple time periods; calculating the local deviation between the sample sequence and the standard sequence in each time period, and taking the time period with the local deviation greater than the local threshold as the difference segment. In response to the number of difference segments being 1, taking the timestamp of the difference segment and the difference sequence between the sample sequence and the standard sequence in the difference segment as a set of training data; in response to the absolute value of the difference between the filling quality and the standard quality being greater than the difference threshold, the noise label of the training data is 0, otherwise, the noise label of the training data is 1; inputting the training data into a binary classification model, outputting the noise probability, and training the binary classification model based on the cross-entropy function between the noise probability and the noise label; inputting the timestamp and the difference sequence of each time period in the real-time sequence into the trained binary classification model, and outputting the noise probability of each time period in the real-time sequence; Weighted summation of the local deviations of each time period according to the noise probability to obtain the global deviation. In response to the global deviation being greater than the global threshold, the monitoring result is abnormal, otherwise the monitoring result is normal.

2. The processing and filling monitoring method based on multi-source sensing data according to claim 1, wherein Sensor data The calculation method for the correlation between and filling quality includes: Take the sensor data The DTW distance between the historical sequences of sensor data other than is used as the clustering distance, and the sample data is clustered to obtain multiple clusters. The sample data includes the historical sequences of sensor data during the historical canning process and the canning quality; Within any clustering cluster, calculate the sensor data of the average sequence of each historical sequence, and use the variance of the DTW distance between each historical sequence of the sensor data and the average sequence as the first variance, and calculate the absolute value of the difference between the first variance and the canned quality variance in the clustering cluster; The absolute value of the difference of each clustering cluster is weighted and summed according to the proportion of the sample data in the clustering cluster, and the correlation between the sensor data and the filling quality is negatively correlated with the weighted sum result.

3. The processing and filling monitoring method based on multi-source sensing data according to claim 2, wherein Sensor data and the correlation with filling quality is as follows: , is an arbitrary minimum value, is the weighted sum result of the sensor data .

4. The processing and filling monitoring method based on multi-source sensing data according to claim 2, wherein Sensor data The weighted summation result Satisfies the relational expression: ; is the total number of clustering clusters, is the number of sample data in the clustering cluster , is the number of all sample data, and are the first variance and the canned quality variance of the clustering cluster respectively.

5. The processing and filling monitoring method based on multi-source sensing data according to claim 1, characterized in that Moment of the fused data is as follows: ; is the type of sensor data, is the sensor data correlation, is the sum of the correlations of all sensor data, is the moment sensor data dimensionless value.

6. The processing and filling monitoring method based on multi-source sensing data according to claim 1, wherein Before obtaining the real-time sequence of the fused data, the monitoring method further includes: Performing denoising processing on the sensor data using Gaussian filtering.

7. The processing and filling monitoring method based on multi-source sensing data according to claim 1, wherein The method for obtaining the standard sequence includes: Taking the sample sequence with the filling quality greater than the quality threshold as the candidate sequence, and calculating the average value at any moment in all candidate sequences to obtain the standard sequence; The standard quality is the average filling quality of each candidate sequence.

8. The processing and filling monitoring method based on multi-source sensing data according to claim 1, characterized in that Global deviation Satisfy the relation: , is the number of time periods, is the noise probability of the time period in the real-time sequence, is the local deviation of the time period ​ 9. The processing and filling monitoring method based on multi-source sensing data according to claim 1, characterized in that, Training the binary classification model based on the cross-entropy function between the noise probability and the noise label includes: Calculating the proportion of the number of training data in each time period in the total amount of training data, and taking the difference between 1 and the proportion as the normalized weight of the corresponding time period; Calculating the cross-entropy function between the noise probability and the noise label, and training the binary classification model based on the product of the cross-entropy function and the normalized weight of the time period in the training data.

10. The processing and filling monitoring method based on multi-source sensing data according to claim 1, wherein The binary classification model is a fully connected neural network or a logistic classification model.

Images