Plastic manufacturing process anomaly detection method and system based on sensor data fusion

Abstract

The invention provides a plastic manufacturing process anomaly detection method and system based on sensor data fusion, and relates to the technical field of production process data analysis, and the method comprises the following steps: arranging a plurality of sensors on a plastic manufacturing production line; a plurality of real-time detection values are periodically collected based on a plurality of sensors; periodically obtaining a sampling detection result of the plastic finished product; performing first data fusion analysis based on the current real-time detection numerical value and the sampling detection result, and judging whether equipment abnormity exists in the plastic manufacturing production line or not; if the judgment result is that the equipment is abnormal, forming a real-time detection numerical value array and a sampling detection result array according to a time sequence; and performing second data fusion analysis based on the real-time detection numerical value array and the sampling detection result array. The method has the advantages that instantaneous data disturbance is overcome, and then high-reliability identification and traceability positioning of the abnormal state are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of production process data analysis, in particular to a plastic manufacturing process anomaly detection method and system based on sensor data fusion. BACKGROUND

[0002] In the plastic manufacturing process, in order to ensure product performance and consistency, abnormal monitoring is usually performed at each link in the production line.

[0003] A variety of sensors are usually arranged in the production line to collect relevant process parameters for anomaly detection, but the existing anomaly detection mainly relies on manual experience or rule-based threshold judgment. Based on the existing technical scheme, the anomaly detection depending on sensor data is easily affected by environmental fluctuations, noise, drift or short-term disturbance, causing instantaneous data disturbance, and there is a risk of missing detection, resulting in frequent false positives or false negatives, making it difficult to achieve early warning and accurate positioning of equipment anomalies.

[0004] Therefore, it is urgent to improve the plastic manufacturing process anomaly detection method, overcome instantaneous data disturbance, and realize high-reliability identification and traceability positioning of abnormal states, so as to improve the intelligence and stability of plastic manufacturing process detection. SUMMARY

[0005] The purpose of the present application is to provide a plastic manufacturing process anomaly detection method and system based on sensor data fusion, which can overcome instantaneous data disturbance and realize high-reliability identification and traceability positioning of abnormal states.

[0006] The present application is realized by the following technical scheme: The plastic manufacturing process anomaly detection method based on sensor data fusion comprises the following steps: Arranging a plurality of sensors in the plastic manufacturing production line; Periodically collecting a plurality of real-time detection values based on the plurality of sensors respectively; Periodically obtaining sampling detection results of plastic finished products; Performing first data fusion analysis based on the current real-time detection values and the sampling detection results, the first data fusion analysis being used to eliminate the influence of detection errors through data fusion of unqualified sample density evaluation parameters and the real-time detection values, and to judge whether there is equipment anomaly in the plastic manufacturing production line; If the judgment result is that there is no equipment anomaly, no operation is performed; If the judgment result is that there is equipment anomaly, the real-time detection values and the sampling detection results of N detection periods up to the present are obtained, and real-time detection value arrays and sampling detection result arrays are formed in time sequence respectively; A second data fusion analysis is performed based on the real-time detection value array and the sampling detection result array to locate the abnormal part.

[0007] Preferably, the method of arranging multiple sensors in a plastic manufacturing production line is: A temperature sensor is arranged in the material heating area, the inner cavity of the plastic manufacturing mold and the cooling part of the equipment respectively; Arrange a pressure sensor at the front end of the material injection component; Current sensors and voltage sensors are arranged on the circuit connection lines of the manufacturing equipment.

[0008] Preferably, the method for performing the first data fusion analysis based on the real-time detection value and the sampling detection result is: Obtaining an unqualified sample density evaluation parameter based on the sampling test result, wherein the unqualified sample density evaluation parameter is used to comprehensively describe the proportion of unqualified samples in the sampling test result and the degree of parameter deviation; The first data fusion analysis is performed in combination with the unqualified sample density evaluation parameter and the real-time detection value.

[0009] Preferably, the method for obtaining the density evaluation parameter of unqualified samples based on the sampling test results is: Measure a total of J performance parameters of each sample tested separately. If any performance parameter of a sample tested exceeds the corresponding threshold range, it will be judged as an unqualified sample; Obtain the percentage of unqualified samples in the sampling test results : ; in, is the total amount of samples sampled for testing, The number of unqualified samples tested by random sampling; According to the degree to which the performance parameters of the sampled samples deviate from the corresponding threshold range, the parameter deviation degree is obtained : ; ; ; in, The offset value of the jth performance parameter of the mth sampling test sample The penalty function value of is the jth performance parameter threshold range, is the value of the jth performance parameter of the mth sample tested, if represents the condition in the conditional function, is the natural exponential function, a tolerance of the offset and a value range of the offset is ; The unqualified sample density evaluation parameter is obtained by fusing the unqualified sample proportion in the sampling detection result and the offset degree of the parameter : ; wherein, is an offset weight adjustment factor, is a nonlinear amplification index, and The influence degree of the unqualified sample density evaluation parameter on the parameter offset degree is obtained by fitting training.

[0010] Preferably, the method of the first data fusion analysis is: A fusion evaluation function is established based on the unqualified sample density evaluation parameter and the number of abnormal values in the real-time detection value, and a fusion score value R is obtained by the fusion evaluation function: ; wherein, is an adjustment constant, represents the value of the real-time detection data of the s-th sensor, represents the value threshold range of the s-th sensor, represents the number of sensors, is a true function; When the value of the fusion score value R is greater than a preset score threshold, it is judged that there is a device abnormality, otherwise it is judged that there is no device abnormality.

[0011] Preferably, the method of the second data fusion analysis is: All samples with qualified sampling detection results are removed from the sampling detection result array to form a sampling unqualified array in time sequence; Adjustment coefficient correlation analysis is performed according to the real-time detection value array and the sampling unqualified array to locate the sensors with abnormal values; Abnormal analysis is performed on the sensors with abnormal values to determine whether the corresponding point position is abnormal; The adjustment coefficient of the adjustment coefficient correlation analysis is updated according to the result of the second data fusion analysis.

[0012] Preferably, the method of adjustment coefficient correlation analysis according to the real-time detection value array and the sampling unqualified array is: The types of performance parameters corresponding to all unqualified performance parameters appearing in the unqualified array are obtained, denoted as unqualified performance parameter types; ​obtaining the average value of each detection cycle of each unqualified performance parameter type in the unqualified array , representing the average value of the u-th unqualified performance parameter type in the unqualified array in the v-th detection cycle, , , representing the total number of unqualified performance parameter types; performing the assigned adjustment coefficient correlation analysis according to the average value of the u-th unqualified performance parameter type in each detection cycle and the value of the s-th sensor in the real-time detection value array, , , representing the number of sensors, and the method of the assigned adjustment coefficient correlation analysis is: ; ; ; wherein, the evaluation parameter of the assigned adjustment coefficient correlation analysis, and is an array intermediate parameter, is a function for calculating the correlation between two arrays, representing the real-time detection value of the s-th sensor in the n-th detection cycle, n = 1, 2, …, N, is the initial value of the adjustment coefficient when there is no feedback adjustment and is 1; the evaluation parameter of the assigned adjustment coefficient correlation analysis is greater than a preset evaluation parameter threshold then determining that the s-th sensor is a sensor with abnormal value.

[0013] Preferably, the method for performing abnormal analysis on the sensor with abnormal value is to perform the following operations respectively for each sensor determined to have abnormal value: assuming that the sensor determined to have abnormal value is the k-th sensor, ; obtaining all the evaluation parameters of the assigned adjustment coefficient correlation analysis of the k-th sensor , ; obtaining the abnormal score of the k-th sensor based on the evaluation parameters of the assigned adjustment coefficient correlation analysis : ; ; wherein, a function representing maximum value of u-based change, is an amplification coefficient, and are the average value and standard deviation of the kth sensor in the real-time detection value array respectively, if(·) is a true value function; if the value of is greater than a preset abnormal score threshold threshold , it is determined that the target detected by the kth sensor is an abnormal part.

[0014] Preferably, the method for adjusting the weight of the correlation analysis of the adjustment coefficient according to the result of the second data fusion analysis is: a sensor influence trajectory diagram is established, the sensor influence trajectory diagram includes a plurality of sensor nodes and a plurality of performance parameter nodes, the sth sensor node represents the sth sensor respectively, the jth performance parameter node represents the jth performance parameter of the plastic respectively, , , J represents the number of performance parameters; each time the abnormal part is located, if the result is that the target detected by the kth sensor is an abnormal part, a vector from the kth sensor node to the dth performance parameter node is added, and the vector is assigned a weight of 1, the dth performance parameter is the performance parameter related to the abnormal value of the kth sensor in the first data fusion analysis, , ; after the second data fusion analysis is performed, the detection part of the kth sensor is repaired, if the improvement of the dth performance parameter after the repair is less than a threshold, the weight of the vector from the kth sensor node to the dth performance parameter node is modified to , , otherwise no operation is performed; when the correlation analysis of the adjustment coefficient is performed, if the correlation analysis of the adjustment coefficient is performed according to the average value of each detection cycle of the fth unqualified performance parameter type and the value of the kth sensor in the real-time detection value array, the performance parameter node corresponding to the fth unqualified performance parameter type and the kth sensor node are obtained, and the weight of all vectors between the nodes is obtained, ; the sensor influence trajectory diagram is checked, and the corresponding adjustment coefficient is updated : ; wherein, is a positive number to avoid a denominator of 0, is the weight of the yth vector between the nodes, and Y is the total number of vectors between the nodes.

[0015] The application also provides a plastic manufacturing process anomaly detection system based on sensor data fusion, applied to the plastic manufacturing process anomaly detection method based on sensor data fusion, and characterized in that it comprises: a sensor module, in which multiple sensors are arranged on a plastic manufacturing production line; a data detection acquisition module, configured to periodically collect multiple real-time detection values based on the multiple sensors respectively; a sampling data acquisition module, configured to periodically acquire sampling detection results of plastic finished products; a first data fusion analysis module, configured to perform first data fusion analysis based on the real-time detection values and the sampling detection results, wherein the first data fusion analysis is used to eliminate the influence of detection errors through data fusion of unqualified sample density evaluation parameters and the real-time detection values, and to determine whether there is equipment anomaly on the plastic manufacturing production line; a time series array acquisition module, configured to, when the determination result is that there is equipment anomaly, acquire the real-time detection values and the sampling detection results of N detection cycles before the current time, and form real-time detection value arrays and sampling detection result arrays in time sequence respectively; a second data fusion analysis module, configured to perform second data fusion analysis based on the real-time detection value arrays and the sampling detection result arrays, and locate the position of the anomaly.

[0016] The technical scheme of the application has at least the following advantages and beneficial effects: The application combines the multiple sensor real-time detection values in the plastic production process with the sampling detection results, realizes fusion analysis of real-time detection and finished product quality detection, effectively suppresses inaccurate judgment caused by instantaneous data errors or fluctuations under a single data source, and improves the accuracy and robustness of anomaly detection; In the first data fusion, the application comprehensively considers the proportion of unqualified samples in sampling detection and the offset degree of various performance parameters, as well as the sensor values of each key point on the production line, constructs a more comprehensive measurement index for the plastic production state, can more comprehensively reflect the process stability and production state, and improves the sensitivity and discrimination accuracy of the process anomaly; In the first data fusion, the application determines whether there is anomaly through data of a single point in time, can perform high-speed analysis based on simple data volume, rapidly responds at a time point where there may be anomaly, and then performs more in-depth second data fusion analysis based on multi-cycle data for the extracted suspected anomaly point, realizes the tracing and positioning of the anomaly position, and further improves the reliability of anomaly detection; The application has reasonable design, convenient deployment, is suitable for various plastic product manufacturing scenes, and has good engineering adaptability and promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart of a sensor data fusion-based plastic manufacturing process anomaly detection method provided for Embodiment 1 of the present application is shown in FIG. 1. Figure 2 A principle diagram of a sensor data fusion-based plastic manufacturing process anomaly detection system provided for Embodiment 2 of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0019] Embodiment 1 The present embodiment provides a sensor data fusion-based plastic manufacturing process anomaly detection method, referring to FIG. 1, which comprises the following steps: Figure 1 Step S01: arranging multiple sensors on a plastic manufacturing production line. In the present embodiment, the method of arranging multiple sensors on a plastic manufacturing production line is as follows: arranging one temperature sensor in each of a material heating zone, a plastic manufacturing mold cavity, and a cooling component of the equipment; arranging a pressure sensor at the front end of a material injection component; arranging a current sensor and a voltage sensor on the circuit connection line of the manufacturing equipment.

[0020] Step S02: periodically collecting multiple real-time detection values based on the multiple sensors respectively, and periodically obtaining sampling detection results of plastic finished products.

[0021] It is particularly pointed out that, when the sampling detection results of plastic finished products are periodically obtained, the sampling detection object extracted is the plastic finished products between the last detection cycle and the present detection cycle.

[0022] When sampling detection is performed, multiple performance parameters can be evaluated, such as dimensional accuracy, tensile strength, surface roughness detected by a profilometer, color consistency detected by a color difference meter, etc.

[0023] Step S03: performing first data fusion analysis based on the real-time detection values and the sampling detection results, wherein the first data fusion analysis is used to eliminate the influence of detection errors through data fusion of unqualified sample density evaluation parameters and the real-time detection values, and to determine whether there is equipment anomaly in the plastic manufacturing production line. ​

[0024] As a preferred solution, the method for performing the first data fusion analysis based on the real-time detection value and the sampling detection result is: an unqualified sample density evaluation parameter is obtained based on the sampling detection result, the unqualified sample density evaluation parameter being used to comprehensively describe the unqualified sample proportion and the parameter deviation degree of the sampling detection result; the first data fusion analysis is performed in combination with the unqualified sample density evaluation parameter and the real-time detection value.

[0025] Specifically, the method for obtaining the unqualified sample density evaluation parameter based on the sampling detection result is: a total of J performance parameters of each sampling detection sample are measured, and if any performance parameter of a sampling detection sample exceeds the corresponding threshold range, the sampling detection sample is judged as an unqualified sample; an unqualified sample proportion of the sampling detection result is obtained : ; wherein, is the total number of sampling detection samples, is the number of unqualified sampling detection samples; a parameter deviation degree is obtained according to the degree of deviation of the performance parameter of the sampling detection sample from the corresponding threshold range : ; ; ; wherein, is the deviation value of the jth performance parameter of the mth sampling detection sample is the penalty function value of is the threshold range of the jth performance parameter, is the value of the jth performance parameter of the mth sampling detection sample, if represents the condition in the conditional function, is a natural exponential function, and according to experience is the deviation tolerance and its value range is ; the unqualified sample density evaluation parameter is obtained by fusing the unqualified sample proportion of the sampling detection result and the parameter deviation degree : ; wherein, is a deviation weight adjustment factor, is a nonlinear amplification index, and for controlling the degree of parameter deviation The degree of influence of the unqualified sample density evaluation parameter is evaluated and obtained by fitting training.

[0026] On this basis, the method of the first data fusion analysis is: A fusion evaluation function is established based on the unqualified sample density evaluation parameter and the number of abnormal values in the real-time detection values, and a fusion score value R is obtained by the fusion evaluation function: ; Wherein, is an adjustment constant, represents the value of the real-time detection data of the s-th sensor, represents the value threshold range of the s-th sensor, represents the number of sensors, is a true function; When the value of the fusion score value R is greater than a preset score threshold, it is judged that there is a device abnormality, otherwise it is judged that there is no device abnormality.

[0027] In the above scheme, the unqualified sample density evaluation parameter is used to quantify the quality level of sampling detection, and the degree of parameter deviation and the proportion of unqualified samples are obtained together to construct the unqualified sample density evaluation parameter as a comprehensive index, through data analysis of sensors other than sensors, the abnormal data analysis deviation caused by short-term errors of sensors caused by external influences is eliminated. When calculating the unqualified sample density evaluation parameter , a non-linear amplification mechanism is introduced , and the serious deviation is significantly amplified, which improves the sensitivity to systematic serious abnormalities. An adjustable deviation weight adjustment factor is introduced, if the field requires special sensitivity to parameter deviation . Through the adjustable and trained and , the entire system has good flexibility and adaptability. The greater the value of the unqualified sample density evaluation parameter , the worse the comprehensive quality of sampling detection.

[0028] At the same time, in the calculation of the degree of parameter deviation , a penalty function is used, and the degree of deviation of each performance parameter from the threshold value is introduced in the penalty function, which significantly enhances the detection ability of slight abnormal fluctuations, and improves the recognition sensitivity of early faults through data fusion of multiple performance parameters. It represents the offset value of the jth performance parameter of the mth sample, that is, the relative deviation of the performance index from its allowable threshold range. The penalty function is processed, and the penalty function Item, in When the penalty function is particularly small, the value is close to 0. On the contrary, when the penalty function is large, the value of the penalty function will approach 1. Finally, the degree of parameter deviation It is the average deviation degree of all unqualified samples in all test parameters.

[0029] Since the operation of equipment in the plastic manufacturing process has complex nonlinear characteristics, the real-time data of the sensor may have fluctuation errors, and relying solely on real-time data is very likely to cause misjudgment. Abnormalities on the production line will inevitably cause abnormal quality of plastic finished products. Therefore, this implementation calculates the density evaluation parameters of unqualified samples. Finally, the abnormality judgment is made by combining the quality of the sampling test and the detection results of multiple sensors. In other words, this implementation effectively compensates for the lag of sampling detection and the volatility of sensor data through the data fusion mechanism, and has higher robustness and anti-interference ability. Specifically, in the fusion evaluation function to obtain the fusion score value R, the unqualified sample density evaluation parameter is combined The number of sensors that are not within the standard threshold range is evaluated. The more abnormal sensors there are, the The larger the value. The processing makes the number of smooth score growth and also facilitates the adjustment of the influence of the sensor. The purpose of v is to prevent the system from calculating errors when ln0 (that is, all sensors are normal). A larger v will make the parameters inside ln(.) more sensitive. The value of v can also be obtained through fitting experiments.

[0030] If the judgment result of the above scheme is that there is no device abnormality, no operation is performed and the system directly waits for the next detection cycle; If the judgment result of the above solution is that there is a device abnormality, then go to step S04.

[0031] Step S04: obtaining the real-time detection values ​​and the sampling detection results of the N detection cycles up to the current time, and forming a real-time detection value array and a sampling detection result array respectively in time sequence; Step S05: performing a second data fusion analysis based on the real-time detection value array and the sampling detection result array to locate the abnormal part.

[0032] As a preferred solution of this embodiment, the method for performing the second data fusion analysis is: From the sampling test result array, all sampling test result qualified samples are removed, and a sampling unqualified array is formed in time sequence. The purpose of this step is to only keep the problem samples that are helpful for positioning abnormalities.

[0033] According to the real-time detection value array and the sampling unqualified array, a regulation coefficient correlation analysis is performed to locate the sensor with numerical abnormalities; Abnormal analysis is performed on the sensor with numerical abnormalities to determine whether the corresponding point position has abnormalities; According to the results of the second data fusion analysis, feedback is performed and the regulation coefficient of the regulation coefficient correlation analysis is updated.

[0034] The above scheme aims to perform multi-point data analysis on the sensors judged to have abnormalities after the first data fusion analysis, to realize more accurate further judgment to determine whether there are abnormalities and locate the abnormal position. And through the result feedback, the important parameters for positioning the abnormal position are adjusted, and the reliability of the calculation is updated in real time, that is, after the fusion feedback regulation mechanism, the long-term accuracy is guaranteed.

[0035] When locating the sensor with numerical abnormalities, the method for performing regulation coefficient correlation analysis according to the real-time detection value array and the sampling unqualified array is: Get the types of performance parameters corresponding to all unqualified performance parameters in the unqualified array, denoted as unqualified performance parameter types; Get the average value of each detection cycle of each unqualified performance parameter type in the unqualified array represents the average value of the u-th unqualified performance parameter type in the unqualified array in the v-th detection cycle, represents the total number of unqualified performance parameter types; According to the average value of the u-th unqualified performance parameter type in each detection cycle and the value of the s-th sensor in the real-time detection value array, a regulation coefficient correlation analysis is performed, represents the number of sensors, and the method for performing regulation coefficient correlation analysis is: Among them, the evaluation parameters of the regulation coefficient correlation analysis, and ​​​​​​​​an array of intermediate parameters, a function for calculating the correlation between two arrays, represents the real-time detection value of the s-th sensor in the n-th detection cycle, n = 1, 2, …, N, is the initial value of the adjustment coefficient when there is no feedback adjustment, and is 1; evaluation parameters of the correlation analysis with the adjustment coefficient greater than a preset evaluation parameter threshold Then, the s-th sensor is determined to be a sensor with abnormal values.

[0036] Next, the method for performing abnormal analysis on the sensor with abnormal values is as follows: Let the sensor determined to have abnormal values be the k-th sensor, ; obtain all the evaluation parameters of the correlation analysis with the adjustment coefficient of the k-th sensor , ; obtain the abnormal score of the k-th sensor based on the evaluation parameters of the correlation analysis with the adjustment coefficient : ; ; wherein, represents a function for calculating the maximum value based on the change of u, is an amplification coefficient, and are the average value and the standard deviation of the k-th sensor in the array of real-time detection values, respectively, and if( ) is a true value function; If the value of is greater than a preset abnormal score threshold , the target detected by the k-th sensor is determined to be an abnormal part.

[0037] In addition, the method for performing feedback adjustment on the weight of the correlation analysis with the adjustment coefficient based on the result of the second data fusion analysis is preferably as follows: establish a sensor influence trajectory diagram, the sensor influence trajectory diagram includes a plurality of sensor nodes and a plurality of performance parameter nodes, the s-th sensor node represents the s-th sensor, and the j-th performance parameter node represents the j-th performance parameter of the plastic, , , and J represents the number of performance parameters; If the result is that the target detected by the kth sensor is the abnormal part, a vector is added from the kth sensor node to the dth performance parameter node, and the vector is assigned a weight of 1, and the dth performance parameter is the performance parameter related to the abnormal value of the kth sensor in the first data fusion analysis, , ; After the second data fusion analysis is performed, the detection part of the kth sensor is repaired, and if the improvement of the dth performance parameter after the repair is less than a threshold value, the weight of the vector from the kth sensor node to the dth performance parameter node is modified to , , otherwise no operation is performed; When the adjustment coefficient correlation analysis is performed, if the adjustment coefficient correlation analysis is performed according to the average value of each detection cycle and the value of the kth sensor in the real-time detection value array of the fth unqualified performance parameter type, the performance parameter node corresponding to the fth unqualified performance parameter type and the kth sensor node are obtained, and the weight of all vectors between the nodes is obtained, ; The sensor influence trajectory diagram is checked, and the corresponding adjustment coefficient is updated : ; wherein, is a positive number to avoid a denominator of 0, is the weight of the yth vector between the nodes, and Y is the total number of vectors between the nodes.

[0038] In the above scheme, first, for the abnormal performance parameter, the adjustment coefficient correlation analysis is performed, and the data analysis is deduced in reverse with the unqualified performance parameter type as the basis to locate the abnormal point. In the adjustment coefficient correlation analysis, the correlation of the time series is used to find the sensor value with high correlation with the unqualified performance parameter, that is, to locate the suspected abnormal sensor, and to further analyze whether there is a problem with such suspected abnormal sensor value. It is particularly noted that when the adjustment coefficient correlation analysis is performed, the unqualified data of the sample is aggregated into a time series, which is helpful for alignment analysis with real-time sensor data, and The r(.,.) in the above formula can be calculated using the Pearson correlation coefficient.

[0039] Next, the abnormal score can be calculated to further confirm the abnormal sensor. The abnormal score The correlation and sensor volatility are fused in the calculation, that is, the evaluation parameter of correlation analysis of the adjustable coefficient is directly introduced, and the ratio of the standard deviation and the average value of the sensor in the real-time detection value array is detected in real time. If the sensor abnormal data fluctuation is large and highly correlated with the unqualified result, the score will naturally increase. The correlation value of the performance parameter with the largest impact is directly taken as the judgment, that is, the sensor fluctuation is large and affects any performance parameter, which is judged as an abnormal point.

[0040] The adjustable coefficient of the embodiment supports the subsequent feedback mechanism and improves the adaptability of the method. The feedback rule is mainly that if the performance improves little after maintenance, the edge weight is adjusted to be low, and the weight change will affect the value of the adjustable coefficient in the correlation analysis of the next round. And here the graph model is added to make the correlation weight have historical memory. This feedback mechanism effectively avoids continuous misleading due to one or two misjudgments, and improves the judgment accuracy.

[0041] Embodiment 2 The application also provides a plastic manufacturing process anomaly detection system based on sensor data fusion, referring to Figure 2 , applied to the plastic manufacturing process anomaly detection method based on sensor data fusion, characterized in that it comprises: A sensor module is arranged with multiple sensors on a plastic manufacturing production line. A data detection acquisition module is used to periodically collect multiple real-time detection values based on multiple sensors. A sampling data acquisition module is used to periodically acquire sampling detection results of plastic finished products. A first data fusion analysis module is used to perform first data fusion analysis based on the current real-time detection values and the sampling detection results. The first data fusion analysis is used to eliminate the influence of detection errors through the data fusion of the unqualified sample density evaluation parameter and the real-time detection values, and to judge whether there is equipment anomaly in the plastic manufacturing production line. A time series array acquisition module is used to acquire the real-time detection values and the sampling detection results of N detection cycles up to the present when the judgment result is that there is equipment anomaly, and to form real-time detection value arrays and sampling detection result arrays in time series, respectively. A second data fusion analysis module is used to perform second data fusion analysis based on the real-time detection value array and the sampling detection result array to locate the abnormal position.

[0042] ​The above merely describes the preferred embodiments of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting anomalies in plastic manufacturing processes based on sensor data fusion, characterized in that: The following steps are involved: Deploy various sensors in plastic manufacturing production lines; Periodically collect multiple real-time detection values ​​based on multiple sensors; Periodically obtain sampling test results of finished plastic products; performing a first data fusion analysis based on the current real-time detection value and the sampling detection result, wherein the first data fusion analysis is used to eliminate the influence of detection errors by fusing the data of the unqualified sample density evaluation parameter and the real-time detection value, and to determine whether there is any equipment abnormality in the plastic manufacturing production line; If the judgment result is that there is no device abnormality, no operation is performed; If the judgment result is that there is a device abnormality, the real-time detection value and the sampling detection result of the N detection cycles up to the current are obtained, and a real-time detection value array and a sampling detection result array are formed in time sequence respectively; A second data fusion analysis is performed based on the real-time detection value array and the sampling detection result array to locate the abnormal part.

2. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 1, characterized in that: The method for arranging multiple sensors in a plastic manufacturing production line is as follows: A temperature sensor is arranged in the material heating area, the inner cavity of the plastic manufacturing mold and the cooling part of the equipment respectively; Arrange a pressure sensor at the front end of the material injection component; Current sensors and voltage sensors are arranged on the circuit connection lines of the manufacturing equipment.

3. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 1, characterized in that: The method for performing the first data fusion analysis based on the real-time detection value and the sampling detection result is: Obtaining an unqualified sample density evaluation parameter based on the sampling test result, wherein the unqualified sample density evaluation parameter is used to comprehensively describe the proportion of unqualified samples in the sampling test result and the degree of parameter deviation; The first data fusion analysis is performed in combination with the unqualified sample density evaluation parameter and the real-time detection value.

4. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 3, characterized in that: The method for obtaining the density evaluation parameter of unqualified samples based on the sampling test results is: Measure a total of J performance parameters of each sample tested separately. If any performance parameter of a sample tested exceeds the corresponding threshold range, it will be judged as an unqualified sample; Obtain the percentage of unqualified samples in the sampling test results : ； in, is the total amount of samples sampled for testing, The number of unqualified samples tested by random sampling; According to the degree to which the performance parameters of the sampled samples deviate from the corresponding threshold range, the parameter deviation degree is obtained : ； ； ； in, The offset value of the jth performance parameter of the mth sampling test sample The penalty function value of is the jth performance parameter threshold range, is the value of the jth performance parameter of the mth sample tested, if represents the condition in the conditional function, is the natural exponential function, is the offset tolerance and its value range is ; The unqualified sample density evaluation parameter is obtained by integrating the unqualified sample ratio of the sampling test results and the parameter deviation degree. : ； in, is the bias weight adjustment factor, is the nonlinear amplification exponent, and Used to control the degree of parameter deviation The degree of influence on the density evaluation parameters of the unqualified samples is obtained through fitting training.

5. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 4, characterized in that: The first data fusion analysis method is: A fusion evaluation function is established based on the unqualified sample density evaluation parameter and the number of outliers in the real-time detection value, and a fusion score value R is obtained through the fusion evaluation function: ； in, is the adjustment constant, Represents the value of the real-time detection data of the sth sensor, represents the numerical threshold range of the sth sensor, Represents the number of sensors, is a truth function; When the value of the fusion score value R is greater than a preset score threshold, it is determined that there is a device abnormality; otherwise, it is determined that there is no device abnormality.

6. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 1, characterized in that: The method for performing the second data fusion analysis is: Eliminate all samples with qualified sampling test results from the sampling test result array, and form a sampling unqualified array in time sequence; Performing correlation analysis on the adjustment coefficients based on the real-time detection value array and the sampling unqualified array to locate the sensor with abnormal values; Perform abnormal analysis on sensors with abnormal values ​​to determine whether the corresponding points have abnormalities; Feedback is provided based on the result of the second data fusion analysis to update the adjustment coefficient of the adjustment coefficient correlation analysis.

7. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 6, characterized in that: The method for performing correlation analysis on the adjustment coefficients according to the real-time detection value array and the sampling unqualified array is as follows: Obtaining the types of performance parameters corresponding to all unqualified performance parameters appearing in the unqualified array, and recording them as unqualified performance parameter types; Get the average value of each unqualified performance parameter type in each detection cycle in the unqualified array , represents the average value of the u-th unqualified performance parameter type in the unqualified array in the v-th detection cycle, , , Represents the total number of unqualified performance parameter types; According to the average value of the u-th unqualified performance parameter type in each detection cycle and the value of the s-th sensor in the real-time detection value array, the adjustment coefficient correlation analysis is performed. , , represents the number of sensors, and the method for correlation analysis of the adjustment coefficient is: ； ； ； in, The evaluation parameters of the adjustment coefficient correlation analysis are: and is the intermediate parameter of the array, To find the correlation function between two arrays, Represents the real-time detection value of the sth sensor in the nth detection cycle, n=1,2,…,N, The initial value of the adjustment coefficient without any feedback adjustment is 1; Evaluation parameters of the adjustment coefficient correlation analysis Greater than the preset evaluation parameter threshold The sth sensor is judged to be the sensor with abnormal value.

8. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 7, characterized in that: The method for performing abnormal analysis on sensors with abnormal values ​​is to perform the following operations on each sensor that is determined to have abnormal values: Assume that the sensor that is judged to have a numerical abnormality is the kth sensor, ; Obtain the evaluation parameters of the correlation analysis of all the adjustment coefficients of the kth sensor , ; Obtain the abnormality score of the kth sensor based on the evaluation parameters of the adjustment coefficient correlation analysis : ； ； in, represents the function that finds the maximum value based on the change of u, is the magnification factor, and are the mean value and standard deviation of the kth sensor in the real-time detection value array, respectively, and if(.) is a truth function; like The value is greater than the preset abnormality score threshold , then the target detected by the kth sensor is judged to be an abnormal part.

9. The method for detecting anomalies in a plastic manufacturing process based on sensor data fusion according to claim 8, characterized in that: The method for performing feedback adjustment on the weight of the adjustment coefficient correlation analysis according to the result of the second data fusion analysis is: A sensor impact trajectory diagram is established, wherein the sensor impact trajectory diagram includes multiple sensor nodes and multiple performance parameter nodes, wherein the s-th sensor node represents the s-th sensor, and the j-th performance parameter node represents the j-th performance parameter of the plastic. , , J represents the number of performance parameters; Each time the positioning of the abnormal part is performed, if the result is that the target detected by the kth sensor is the abnormal part, a vector pointing from the kth sensor node to the dth performance parameter node is added, and the vector is assigned a weight of 1. The dth performance parameter is the performance parameter related to the abnormal value of the kth sensor in the first data fusion analysis. , ; After performing the second data fusion analysis, the detection part of the kth sensor is repaired. If the improvement of the dth performance parameter after the repair is less than the threshold, the weight of the vector pointing from the kth sensor node to the dth performance parameter node is modified to , , otherwise no operation is performed; When performing the adjustment coefficient correlation analysis, if the adjustment coefficient correlation analysis is performed based on the average value of the f-th unqualified performance parameter type in each detection cycle and the value of the k-th sensor in the real-time detection value array, the performance parameter node and the k-th sensor node corresponding to the f-th unqualified performance parameter type are obtained, and the weights of all vectors between the nodes are obtained. ; Check the sensor impact trajectory map and update the corresponding adjustment coefficient : ； in, To avoid positive numbers with denominators of 0, is the weight of the yth vector between nodes, and Y is the total number of vectors between nodes.

10. A plastic manufacturing process anomaly detection system based on sensor data fusion, applied to the plastic manufacturing process anomaly detection method based on sensor data fusion according to any one of claims 1 to 9, characterized in that: include: Sensor modules, which deploy multiple sensors in plastic manufacturing production lines; A data detection and acquisition module is used to periodically collect multiple real-time detection values ​​based on multiple sensors; Sampling data acquisition module, used to periodically obtain sampling test results of finished plastic products; a first data fusion analysis module, configured to perform a first data fusion analysis based on the current real-time detection value and the sampling detection result, wherein the first data fusion analysis is configured to eliminate the influence of detection errors by fusing the data of the unqualified sample density evaluation parameter and the real-time detection value, and to determine whether there is any equipment abnormality in the plastic manufacturing production line; A time series array acquisition module is used to obtain the real-time detection values ​​and the sampling detection results of the N detection cycles up to the current time when the judgment result is that the device is abnormal, and respectively form a real-time detection value array and a sampling detection result array in time sequence; The second data fusion analysis module is used to perform a second data fusion analysis based on the real-time detection value array and the sampling detection result array to locate the abnormal part.

Images