A working condition identification method

Abstract

This invention discloses a working condition identification method, comprising the following steps: aligning the feature sets of the working conditions based on their time delay; normalizing all datasets and calculating the similarity s between corresponding features X and Y in two sets of time-series data; and calculating the similarity s between two sets of time-series datasets for the same working condition by calculating the similarity s of each corresponding feature. i (i = 1, 2, ..., m), where m represents the number of features of the working conditions; calculate the overall similarity S; calculate the probability P of each group of normal samples participating in the model construction. j The similarity S among all normal states is obtained by calculating the similarity between the dataset of normal equipment states and the dataset of case states using the normal operation data model. normal Similarity S with all case states abnormal By combining the similarity values ​​of the two states to calculate the early warning baseline L, this invention can extend the evaluation method to the process industry to identify working conditions. By using a case library to improve the selection of the early warning baseline, it greatly improves the situation of missed reports and can also adjust the model to effectively control the false alarm rate.

Description

Technical Field

[0001] This invention relates to a modeling method for intelligent anomaly assessment models, and in particular to a working condition identification method. Background Technology

[0002] Patent CN102270271 B, "Method and System for Early Warning and Optimization of Equipment Faults Based on Similarity Curves," describes a warning method that uses a set of typical characteristic data reflecting the actual operating patterns of equipment as the data analysis object, and selects a normal dataset from the data to construct a normal operation data model. The minimum similarity value in the normal dataset serves as the warning baseline, and when the model is applied, the real-time status is evaluated based on this baseline to determine if it is abnormal.

[0003] When applying patent CN102270271 B: "Method and System for Early Warning and Optimization of Equipment Failure Based on Similarity Curves" to the process industry, the following problems arose:

[0004] First, this method uses discrete datasets, which cannot assess trend anomalies.

[0005] Second, this method uses a normal data model built from a dataset in normal operation, and the warning baseline also uses the similarity in normal data, which can easily lead to missed reports.

[0006] Third, in this method, all variable parameters of the device are sampled at the same time in each sampled data group, without data alignment, so the feature values ​​displayed during the warning are not representative.

[0007] Analysis of the causes of the problems in the above applications:

[0008] First, this modeling method does not support using a time series dataset as a sample; when the data size is normal but the trend is abnormal, the model cannot identify this anomaly.

[0009] Second, because the model does not consider the similarity of case data, it uses the minimum similarity among normal data as the warning baseline. Compared to normal data, the similarity of abnormal case data may be higher than that of some normal data, leading to missed cases.

[0010] Third, feature delays are common in process industries. This method does not take into account the delay characteristics, resulting in a lack of alignment between features. Therefore, the feature data displayed during the warning is insufficient to represent the actual abnormal state.

[0011] To apply this assessment method to the process industry for condition identification, the above-mentioned problems were addressed and the assessment method was optimized and improved. Summary of the Invention

[0012] To address the problems existing in the prior art, this invention provides a working condition identification method that uses time-series data to construct an evaluation model and case data to participate in the calculation of early warning baselines. This evaluation method can be extended to the process industry to identify working conditions. The case library is used to improve the selection of early warning baselines, which greatly improves the situation of missed alarms. The model can also be adjusted to effectively control the false alarm rate.

[0013] The objective of this invention is achieved through the following technical solutions.

[0014] A working condition identification method, comprising the following steps:

[0015] Step 1: Alignment based on the time delay between feature sets of operating conditions: First, identify the feature set that reflects the occurrence of the operating condition. If the feature set of the operating condition is {A, B, C}; determine the sampling interval based on the duration of the operating condition change, and organize the time series dataset of the operating condition from the time-continuous raw data. The duration of each time series is the duration of the operating condition change. If feature A reacts one minute earlier than other features, then the data a of feature A at t-1 will be aligned. t-1 Data b corresponding to features B and C in t t c t ;

[0016] Step 2: Normalize all datasets. Then, following Step 1, sample the normal state and case state data respectively to obtain the normal state time series dataset and the case state time series dataset, and normalize the datasets.

[0017] Step 3: Calculate the similarity s between corresponding features X and Y in the two sets of time series data;

[0018] Step 4: Calculate the similarity s of two time-series datasets under the same operating condition by calculating the similarity s of each corresponding feature. i (i = 1, 2, ..., m), where m represents the number of features of the working conditions, and the similarity S of the population is calculated.

[0019] Step 5: Calculate the probability P of each group of normal samples participating in the model construction. j ;

[0020] Step 6: Calculate the similarity S of all normal states by combining the datasets of normal device states and the datasets of case states using the normal operation data model. normal Similarity S with all case states abnormal ;

[0021] Step 7: Calculate the warning baseline L by combining the similarity values ​​of the two states:

[0022] L = max(min(S) normal ), S abnormal )

[0023] Calculate the false alarm rate and false negative rate based on this baseline. If the false alarm rate is higher than the user's requirements, data cleaning needs to be intensified, and the process should return to step 2 to reorganize the time series sample set.

[0024] In step 3, the similarity s consists of assessing trend similarity s1 and assessing numerical similarity s2, that is:

[0025] s = k1*s1 + k2*s2

[0026] Where k1 + k2 = 1, the value of k2 is determined by the characteristic distribution interval corresponding to the working case, that is, k2 = max(Q A )-min(Q A ), Q A This is the dataset of feature A in the case study library for this working condition; ρ is the correlation coefficient. X = [x1, x2, ..., x] n ], Y = [y1, y2, ..., y n In summary, this evaluates the similarity between two sets of temporal states of a feature from two perspectives.

[0027] The overall similarity S in step 4 is calculated as follows:

[0028]

[0029] Step 5: Calculate p for each set of normal data. j The similarity data Q between (j = 1, 2, ..., n) and other normal state data j = [S1, S2, ..., S i (i≠j, i=1,2,...,n), to ensure a relatively uniform distribution of normal operating data in the constructed model, a value is set as the similarity threshold S. L By calculating Q j Medium below S L The number N j Calculate the probability P of each group of normal samples participating in the model construction. j The formula is Based on the probability of each sample, a small number of data points are randomly selected from all normal data to serve as the normal operating data model.

[0030] If there are no case studies in step 3, then k2 is set to 0.

[0031] Compared with existing technologies, the advantages of this invention are: it can extend the evaluation method to the process industry to identify working conditions, use a case library to improve the selection of early warning baselines, greatly improve the situation of missed reports, and can also adjust the model to effectively control the false alarm rate. Attached Figure Description

[0032] Figure 1 This is a flowchart of the present invention.

[0033] Figure 2 This is a normalized effect diagram of the present invention.

[0034] Figure 3 This is the dataset for the normal operation data model of this invention. Detailed Implementation

[0035] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0036] A working condition identification method, comprising the following steps:

[0037] Step 1: Align based on the time delay between feature sets of operating conditions. First, identify the feature set that reflects the operating condition. If the feature set is {A, B, C}, determine the sampling interval based on the duration of the operating condition change, and organize the time series dataset of the operating condition from the continuous raw data. The length of each time series is the duration of the operating condition change. If feature A reacts one minute earlier than other features, then the data a of feature A at t-1 is aligned. t-1 Data b corresponding to features B and C in t t c t .

[0038] Step 2: Normalize all datasets. The following modeling methods are all applied to the normalized data. Then, following Step 1, normal state time series datasets and case state time series datasets are obtained by sampling the normal state and case state data respectively, and the datasets are normalized.

[0039] Step 3: Calculate the similarity s between corresponding features X and Y in the two sets of time series data. The similarity s consists of trend similarity s1 and numerical similarity s2, i.e.:

[0040] s = k1*s1 + k2*s2

[0041] Where k1 + k2 = 1, the value of k2 is determined by the characteristic distribution interval corresponding to the working case, that is, k2 = max(Q A )-min(Q A ), Q A This is the dataset of feature A in the case study library for this working condition; ρ is the correlation coefficient. X = [x1, x2, ..., x] n ], Y = [y1, y2, ..., y n ].

[0042] In summary, the similarity between two sets of time-series states of a feature is evaluated from two perspectives. Specifically, if there are no case studies, k2 is set to 0.

[0043] Step 4: Following Step 4, calculate the similarity s of the two time-series datasets for the same operating condition, and then calculate the similarity s of each corresponding feature. i (i = 1, 2, ..., m), where m represents the number of features of the working conditions, and the overall similarity S is calculated as follows:

[0044]

[0045] Step 5: Calculate p for each group of normal data based on the above steps. j The similarity data Q between (j = 1, 2, ..., n) and other normal state data j = [S1, S2, ..., S i (i≠j, i=1,2,...,n). To ensure a relatively uniform distribution of normal operating data in the constructed model, a small value is set as the similarity threshold S. L By calculating Q j Medium below S L The number N j Calculate the probability P of each group of normal samples participating in the model construction. j The formula is as follows. A small number of data points are randomly selected from all normal data sets based on the probability of each sample group, serving as the normal operation data model.

[0046]

[0047] Step 6: Calculate the similarity S of all normal states by combining the datasets of normal device states and the datasets of case states using the normal operation data model. normal Similarity S with all case states abnormal .

[0048] Step 7: Calculate the warning baseline L by combining the similarity values ​​of the two states:

[0049] L = max(min(S) normal ), S abnormal )

[0050] Calculate the false alarm rate and false negative rate based on this baseline. If the false alarm rate is higher than the user's requirements, data cleaning needs to be intensified, and the process should return to step 2 to reorganize the time series sample set.

[0051] Example

[0052] Taking a typical working condition, such as a silo arching chamber, as an example, the following will be explained:

[0053] Step 1: Align the feature sets of the operating conditions based on their time delays. The main characteristics of the silo arching event are: rapid increase in silo weight, significant decrease in roller press current, and decrease in roller gap. Therefore, the feature set for this operating condition is {silo weight, roller press current, roller gap}. Based on the duration of the operating condition's changes, a time-series dataset is extracted from the continuous raw data. The reaction time for this operating condition is approximately 5 minutes, and the sampling interval is set to 1 minute. Since the roller press current and roller gap react one minute faster than the silo weight, the data for the characteristic roller press current and roller gap are shifted forward by one minute.

[0054] Step 2: Normalize all datasets. The following modeling methods are applied to the normalized data. Then, following the above steps, sample the normal state and case state data to obtain the normal state time series dataset and the case state time series dataset, respectively. All normalized datasets are shown in the figure below. Based on reaction time, five points are sampled at a time series, as shown in the case data... Figure 2 As shown.

[0055] Step 3: Calculate the similarity of features (steady-flow bin weight, roller press current, roller gap) between any two sets of time series data in all state time series datasets;

[0056] Step 4: Calculate the total similarity between any two sets of time series data in all normal state time series datasets based on feature similarity, as per Step 3. The similarity coefficients determined according to the feature distribution intervals corresponding to the work case are as follows.

[0057] feature k1 k2 Stable flow warehouse heavy 0.68 0.32 Roller press current 0.3 0.7 Roller gap 0.8 0.2

[0058] Step 5: Based on the above steps, calculate the similarity data between each group of normal data and other normal state data. To ensure a relatively uniform distribution of normal data in the constructed model, a small similarity threshold of 0.1 is set. By calculating the number of similarity values ​​below 0.1 in each group, the probability of each group of normal samples participating in the model construction is calculated. Based on the probability of each group of samples, a small number of data points are randomly selected from all normal data as the normal data model (e.g., ...). Figure 3 (As shown).

[0059] Step 6: Calculate the similarity of all normal states and all case states by using the normal operation data model to obtain the similarity of the datasets of normal state and case states.

[0060] Step 7: Calculate the warning baseline L = 0.7 by combining the similarity values ​​of the two states, and the false alarm rate based on this baseline is 0.1% and the false alarm rate is 0.4%.

[0061] The comparison results are shown in the table below. Without using case data, the baseline L = 0.73, the false negative rate was 6%, and the false positive rate was 0.4%.

[0062] Without sacrificing the false positive rate, it significantly reduces the number of cases that are missed.

[0063] contrast baseline underreporting rate False alarm rate Before improvement 0.73 6.0％ 0.4％ Improved 0.70 0.1％ 0.4％

Claims

1. A method for identifying operating conditions, characterized by the following steps: include: Step 1: Align based on the time delay between the feature sets of the operating conditions: First, find the feature set that reflects the occurrence of the operating condition. If the feature set of the operating condition is... Based on the duration of the operating condition changes, the sampling interval is determined, and a time-series dataset of the operating conditions is compiled from the continuous raw data. The duration of each time series is the duration of the operating condition change, and the features are... If the response is one minute earlier than other characteristics, then the characteristic will be... exist Data Corresponding features exist Data ; Step 2: Normalize all datasets. Then, following Step 1, sample the normal state and case state data respectively to obtain the normal state time series dataset and the case state time series dataset, and normalize the datasets. Step 3: Calculate the corresponding features in the two sets of time series data. and similarity ; Step 4: Calculate the similarity between two time-series datasets for the same operating condition by evaluating the similarity of each corresponding feature. , Characteristic numbers representing operating conditions are used to calculate the overall similarity. ; Step 5: Calculate the probability of each group of normal samples participating in model construction. Based on the probability of each sample, a small number of data are randomly selected from all normal data as the normal operation data model; Step 6: Calculate the similarity of all normal states by combining the datasets of normal device states and the datasets of case states using the normal operation data model. Similarity to all case states ; Step 7: Calculate the warning baseline by combining the similarity values ​​of the two states. : Calculate the false alarm rate and false negative rate based on this baseline. If the false alarm rate is higher than the user's requirements, data cleaning needs to be intensified, and the process should return to step 2 to reorganize the time series sample set.

2. The working condition identification method according to claim 1, characterized in that... Similarity in step 3 By assessing trend similarity And assessing numerical similarity Composition, namely: in, , The value is determined by the characteristic distribution interval corresponding to the working condition case, that is... , It is a feature in the case study library for this working condition. The dataset; , It is the correlation coefficient. , , In summary, this evaluates the similarity between two sets of temporal states of a feature from two perspectives.

3. The working condition identification method according to claim 1, characterized in that... Overall similarity in step 4 The calculation is as follows: 。 4. The working condition identification method according to claim 1, characterized in that... Step 5: Calculate the normal data for each group. Similarity data with other normal state data To ensure a relatively uniform distribution of normal data in the constructed model, a set value is used as the similarity threshold. Through calculation medium to low quantity Calculate the probability of each group of normal samples participating in the model construction. The formula is Based on the probability of each sample, a small number of data are randomly selected from all normal data as the normal operation data model.

5. The working condition identification method according to claim 2, characterized in that... If there are no case studies in step 3, then Take 0.

Images