Aero-engine test stand multi-sensor data space-time correlation detection method

Abstract

The application provides an aero-engine test stand multi-sensor data space-time correlation detection method, comprising: selecting a certain specific test condition data, using different baseline modeling algorithms, respectively based on the physical space relationship of different sections of the aero-engine sensor, the physical space relationship of different measuring rakes and different radii on the same section, and the sensor data in a certain period of time, verifying the response data of the rake sensor and judging the abnormal condition of the sensor, correlating and deeply mining the multi-dimensional information of the aero-engine, and providing data support and important guarantee for the safety of the test system and the aircraft. The application can fully realize the verification and abnormal detection of the sensor response data in a short period under different test conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of air engine ground test data validation, and particularly relates to a method for detecting time-space correlation of multi-sensor data of an air engine test stand. BACKGROUND

[0002] With the development of low-cost sensing and communication technology, the expansion of its application on aircraft and the popularity of mobile terminals, the state and performance parameters of online monitoring and recording systems can be realized. These massive data have the characteristics of dynamic, continuous sampling, multi-source, unstructured, large volume, etc., which brings new opportunities for the study of dynamic behavior, safe operation and monitoring of complex systems. However, before these data are analyzed and used, their correctness and integrity need to be ensured, and the effectiveness of sensor data needs to be validated to solve the problem of invalid data caused by data loss, sensor failure or noise interference, which is usually called sensor data validation or effectiveness verification.

[0003] For the scene of air engine ground test, the data generated by the sensor in this scene is typical high-dimensional time series data, which has the natural attribute of "time continuity" on the time axis. From the spatial point of view, due to the sequential correlation of each unit of the air engine, the coupling between the data of each monitoring section is strong, and it also has strong "spatial correlation". Therefore, the "time continuity" and "spatial correlation" naturally possessed by the test data of the air engine test stand provide a basis for sensor data validation and anomaly detection.

[0004] At present, foreign sensor data validation technology has been successfully applied in the field of aerospace and has achieved great benefits. According to reports, the Arnold Engineering Center in the United States has developed a gas turbine engine test data validity validation system, which has been successfully applied in the development process of F414 (the engine of F / A-18 aircraft of the Navy) and F119 (the engine of F-22 aircraft of the Air Force). The long-term application of data statistical analysis table shows that the application of test data validity validation system assisted experiment can timely find invalid test data, thereby reducing 50% of the post-rework test, saving half of the manpower for data analysis and validity validation, and greatly saving the engine development cost.

[0005] In domestic air engine ground test, although the general automatic test system of the test stand has been successfully developed, and the online state monitoring and fault diagnosis system has been preliminarily developed and applied, the functions are mainly concentrated in the test process, instrument, signal acquisition and test data management, as well as signal upper / lower threshold alarm, test data trend analysis, etc. The effectiveness analysis of test data still mainly depends on manual interpretation, and there is no system for intelligent verification of test data information, and the automatic validation of test data effectiveness has not been truly realized.

[0006] In summary, the prior art lacks an intelligent method for verifying the validity of aero-engine test stand data, which can mine the relevance of test stand test data from the time and space dimensions and accurately and intelligently verify the validity of aero-engine ground test sensor data. SUMMARY

[0007] The technical problem to be solved by the present application is to provide an aero-engine test stand multi-sensor data space-time correlation detection method, which can realize the goals of aero-engine test stand sensor data correlation verification, fault sensor monitoring and identification.

[0008] The method of the present application specifically comprises the following steps:

[0009] Step 1, n sensors are arranged inside the aero-engine, the data of the aero-engine under different test internal working conditions (such as slow vehicle, N2 (high pressure rotor speed): 90%, N2: 95%, high power state, etc.) are identified, and the response readings of each time sequence sampling point of the n sensors under a steady state (if a piece of data has m sampling points in total, the data matrix shape is m*n, that is, the response readings of each sensor at each time sequence sampling point) under one working condition (such as high power state) are extracted as historical data for use in the baseline model training stage (all normal data); one steady state data of a new experiment under any working condition (the high power state is taken as an example in the foregoing, so the high power state is extracted here) is used as detection data in the data verification stage, and the sensor data is divided into three large dimensions of space, physics and time, based on which five sub-dimensions of the same cross-section, the same radius, different rakes, different cross-sections, parameter correlation and original time sequence are divided, which are respectively recorded as sub-dimensions 1, 2, 3, 4 and 5.

[0010] Space dimension: According to the engine sensor measurement section, it can be divided into engine inlet section, fan outlet section, compressor outlet section, low pressure turbine outlet section, bypass inlet section, bypass outlet section. If the sensors are in the same section, they can also be classified according to the same radius and different measuring bars (taking the engine fan outlet (PtM) section as an example, assuming that there are 3 measuring bars (PtM_01, PtM_02, PtM_03) on each measuring bar, and there are 5 sensors on each measuring bar, the classification is as follows: [PtM_01_01, PtM_02_01, PtM_03_01], [PtM_01_01, PtM_02_01, PtM_03_01], [PtM_01_02, PtM_02_02, PtM_03_02], [PtM_01_03, PtM_02_03, PtM_03_03], [PtM_01_04, PtM_02_04, PtM_03_04], [PtM_01_05, PtM_02_05, PtM_03_05]) or the same measuring bar and different radius ([PtM_01_01, PtM_01_02, PtM_01_03, PtM_01_04, PtM_01_05], [PtM_02_01, PtM_02_02, PtM_02_03, PtM_02_04, PtM_02_05], [PtM_03_01, PtM_03_02, PtM_03_03, PtM_03_04, PtM_03_05]), PtM_01_0X is the sensor reading of the Xth measuring point on the first measuring bar of the fan outlet section, PtM_02_0X is the sensor reading of the Xth measuring point on the second measuring bar of the fan outlet section, PtM_03_0X is the sensor reading of the Xth measuring point on the third measuring bar of the fan outlet section, and X takes a value of 1-5;

[0011] If the sensors are not in the same section, classify the sensors with sensors in different sections (classify PtM_01_01 with other section sensors);

[0012] Physical: PtM_01_01 as the total pressure of the fan outlet can be classified with a certain sensor that measures the total temperature / total pressure of the engine inlet, because they have a sequential relationship in the physical section of the engine;

[0013] Time: Each sensor is a set of time series data itself;

[0014] Step 2, for sub-dimension 1, baseline model training stage processing and data validation stage processing;

[0015] Step 3, for sub-dimension 2, baseline model training stage processing and data validation stage processing;

[0016] Step 4, for sub-dimension 3, baseline model training stage processing and data validation stage processing are performed;

[0017] Step 5, for sub-dimension 4, baseline model training stage processing and data validation stage processing are performed;

[0018] Step 6, for sub-dimension 5, data validation stage processing is performed;

[0019] Step 7, sub-dimensions 1, 2, 3 results are combined into spatial dimension results, sub-dimensions 4, 5 anomaly detection results are respectively taken as physical and time dimension results, and the anomaly detection results are combined from the spatial, physical, and time three large dimensions, and the sensor data is validated, and the dimensional analysis is performed on the abnormal sensor readings, and the data is reconstructed according to the abnormal dimension.

[0020] In step 2, the baseline model training stage processing includes: for all sampling points and cross sections of the historical data, the following processing is performed: sequentially difference processing is performed on the data of the same cross section, the same radius, and the same sampling point; each difference value replaces each data of the original sampling point, and finally all sequential difference values process the data of the same sampling point, the same cross section, and the same radius, and all the data are input into a box plot (the box plot is prior art, the upper and lower bottoms of the box are the upper quartile (Q3) and the lower quartile (Q1) of the data respectively, which means that the box contains 50% of the data. Therefore, the height of the box reflects the fluctuation degree of the data to a certain extent. The upper and lower edges represent the maximum and minimum values of the data. Sometimes there are some points outside the box, which can be understood as “abnormal values” in the data), which is called baseline model 1, and the upper and lower limits of different weeks of the cross section based on the historical data are obtained;

[0021] The data validation stage processing includes: selecting a stable state response reading of a sensor A in the detection data (the general order is according to the relationship of the engine cross section, if the same cross section is taken from the flow field center to the outside according to the rake number, the sensor), sequentially difference processing is performed on the data of the same cross section, the same radius, and the same sampling point of the detection data, and whether the sensor is abnormal is judged according to the model (the upper and lower limits of different weeks of the cross section) corresponding to the sensor A in the baseline model 1; when other sensors are validated, the same processing as the data validation stage in step 2 is performed on other sensors in the detection data (such as selecting a sensor B reading of the same cross section and the same radius but different rakes as the sensor A, sequentially difference processing is performed on the data of the same cross section, the same radius, and the same sampling point of the detection data, and whether the sensor is abnormal is judged according to the model (the upper and lower limits of different weeks of the cross section) corresponding to the sensor B in the baseline model 1), and then the abnormal detection results of each sampling point of the n sensors in sub-dimension 1 are output.

[0022] In step 3, the baseline model training stage processing includes: traversing each regressive sensor respectively, randomly dividing training set and test set from historical data (fault-free samples), processing independent variable parameters of the training set (sub-dimension 2 independent variable selection parameter set is the same as the other sensor readings on the same section of the regressive sensor), dependent variable parameter is the regressive sensor data, inputting the random forest model for training, saving the trained random forest model, and obtaining the random forest baseline model of each sensor on sub-dimension 2; establishing a normal distribution of bias value based on historical data, the bias value interval of sensor i is defined as: upper sub_dimension,i is:

[0023]

[0024] wherein represents the mean of the historical bias value of sub-dimension 2 sensor i, represents the standard deviation of the bias value of sub-dimension 2 sensor i;

[0025] lower sub_dimension,i is:

[0026]

[0027] Statistical anomaly detection is performed based on the bias value distribution of each sensor model;

[0028] The data validation stage processing includes: selecting a steady-state response reading of a sensor i, based on the data of the same measuring rake at different radii of the section where the sensor is located, using the sub-dimension 2 baseline model of the sensor to regress and calculate the sensor reading, the bias value of sensor i at the jth sampling point on sub-dimension 2 is calculated by the following formula:

[0029]

[0030] wherein represents the regression value of sensor i at the jth sampling point on sub-dimension 2 using its random forest baseline model, represents the reading of sensor i at the jth sampling point in the detection data;

[0031] Based on different features of the bias value at each moment, the sensor abnormality is judged from the statistical and time sequence angles:

[0032] Statistical angle: input the bias value of sensor i at each moment into the bias value interval of sensor i (lower sub_dimension2,i ,upper sub_dimension2,i ), if the bias value error is within the bias value interval range: The determination result is valid, that is, the jth sampling point of sensor i is 0 in the statistical angle of the second sub-dimension, denoted as The determination result is invalid, that is, the jth sampling point of sensor i is 1 in the statistical angle of the second sub-dimension, denoted as

[0033] Timing angle: the deviation value of each time of sensor i is judged by a certain timing anomaly detection method, such as a statistical process control ewma control chart method, and the formula of the statistical process control ewma control chart method is as follows:

[0034] z t = λx t + (1-λ)z t-1

[0035] wherein the value range of constant λ is 0≤λ≤1, z t is an ewma statistic, x t is the deviation value of the test data point t of sensor i, the range of t is 0≤t≤m, m is the number of experimental sampling points, t is an integer, the initial value z0 of the ewma statistic is the mean value μ0 of the test data of sensor i; x t is an independent random variable, and the variance is σ 2 , the variance t of z is:

[0036]

[0037] The process control ewma control chart method generates the upper limit ucl and the lower limit lcl of each sampling point according to the input data, and the formula is as follows:

[0038]

[0039]

[0040] wherein L is a constant, generally taken as 3, and σ is the standard deviation of the independent random variable x t ; when i increases, lcl and ucl will be stabilized to the following two values:

[0041]

[0042] If the deviation value is between the upper limit and the lower limit of the jth sampling point: The determination result is valid, that is, the jth sampling point of sensor i is 0 in the timing angle of the second sub-dimension, denoted as Otherwise, the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the timing angle of the second sub-dimension, denoted as The or operation is performed on the sensor i results in sub-dimension 2, and the formula is as follows:

[0043]

[0044] The abnormal detection results of each sampling point of the sensor i in sub-dimension 2 are output.

[0045] The same processing as step 3 is performed on other sensors in the detection data.

[0046] In step 3, the same processing as step 3 is performed on other sensors in the detection data, specifically including: selecting a steady-state response reading of a sensor s, based on the data of the same section and different radii of the measuring rake, using the sensor in the sub-dimension 2 baseline model to regress and calculate the sensor reading, and the specific calculation formula of the deviation value of the sensor s at the jth sampling point in the sub-dimension 2 is as follows:

[0047]

[0048] Based on the different characteristics of the deviation value at each moment, the sensor abnormality is judged from the statistical and time sequence angles:

[0049] Statistical angle: the deviation value of the sensor s at each moment is input into the deviation value interval (lower sub_dimension2,i ,upper sub_dimension2,i ) of the sensor s, if the deviation value error is within the deviation value interval range: The determination result is valid, that is, the jth sampling point of the sensor s in the statistical angle of the sub-dimension 2 is 0, and is recorded as Otherwise, the determination result is invalid, that is, the jth sampling point of the sensor s in the statistical angle of the sub-dimension 2 is 1, and is recorded as

[0050] Time sequence angle: the deviation value of the sensor s at each moment is judged by using a certain time sequence anomaly detection method such as statistical process control ewma control chart method, if the deviation value is between the upper limit and the lower limit of the sampling point j: The determination result is valid, that is, the jth sampling point of the sensor s in the time sequence angle of the sub-dimension 2 is 0, and is recorded as Otherwise, the determination result is valid, that is, the jth sampling point of the sensor s in the time sequence angle of the sub-dimension 2 is 1, and is recorded as The or operation is performed on the sensor s results in sub-dimension 2, and the formula is as follows:

[0051]

[0052] Output the abnormal detection results of each sampling point of the sensor i on the sub-dimension 2; finally output the abnormal detection results of each sampling point of the n sensors on the sub-dimension 2.

[0053] In step 4, the baseline model training stage processing includes: traversing each regressionable sensor respectively, randomly dividing training set and test set from historical data (fault-free samples), processing independent variable parameters of the training set (sub-dimension 3 independent variable selection parameter set is the top 10 sensor readings on different sections associated with the regression sensor), dependent variable parameter is the regression sensor data, inputting the random forest model for training, saving the trained random forest model, and obtaining the baseline model of each sensor on the sub-dimension 3; based on historical data, a normal distribution of bias value is established, and the bias value interval of the sensor i is defined as: upper limit upper sub_dimension3,i is:

[0054]

[0055] wherein represents the mean of the historical bias value of the sub-dimension 3 sensor i, represents the standard deviation of the bias value of the sub-dimension 3 sensor i;

[0056] lower limit lower sub_dimension,i is:

[0057]

[0058] Statistical anomaly detection is performed based on the bias value distribution of each sensor model;

[0059] The data validation stage processing includes: selecting a steady-state response reading of a sensor i, based on the top 10 sensor data on different sections associated with the sensor, using the sub-dimension 3 baseline model of the sensor to perform regression calculation on the sensor reading, and the bias value of the sensor i on the sub-dimension 3 at the jth sampling point is The calculation formula is:

[0060]

[0061] wherein represents the regression value of the sensor i on the sub-dimension 3 at the jth sampling point using its random forest baseline model, represents the reading of the sensor i at the jth sampling point in the detection data;

[0062] Based on different features of the bias value at each moment, the sensor abnormality is judged from the statistical and time sequence angles:

[0063] Statistical angle: input the bias value of the sensor i at each moment into the bias value interval (lower sub_dimension3,iupper sub_dimension3,i If the deviation value error is within the deviation value range The result is valid if the j-th sampling point of sensor i is 0 from the statistical perspective of sub-dimension 2, denoted as... Otherwise, the result is invalid, meaning the j-th sampling point of sensor i is 1 from the statistical perspective of sub-dimension 2, denoted as

[0064] From a timing perspective: The deviation value of sensor i at each moment is judged using a timing anomaly detection method such as the statistical process control EWMA control chart method. If the deviation value is at the upper limit of sampling point j... and lower limit between: The result is valid if the j-th sampling point of sensor i is 0 in the temporal dimension 3, denoted as . Otherwise, the result is invalid, meaning the j-th sampling point of sensor i is 1 in the temporal dimension 3, denoted as

[0065] The OR operation is performed on the sensor i result in sub-dimension 3, as shown in the following formula:

[0066]

[0067] Output the anomaly detection result for each sampling point of sensor i in sub-dimension 3;

[0068] For other sensors, perform the same processing as in step 4, the data verification stage, based on the detection data.

[0069] In step 4, the same processing as in step 4 (data verification stage) is performed on other sensors based on the detection data. Specifically, this includes: selecting the steady-state response reading of sensor s; based on the top 10 sensor data with the highest correlation across different cross-sections of the sensor; using the sensor's baseline model in sub-dimension 3 to perform regression calculation on the sensor reading; and determining the deviation value of sensor s at the j-th sampling point in sub-dimension 3. The calculation formula is:

[0070]

[0071] Based on the different characteristics of the deviation value at each time step, the abnormal situation of the sensor is judged from both statistical and temporal perspectives: Statistical perspective: The deviation value of sensor s at each time step is input into the deviation value range of sensor s (lower... sub_dimension,i upper sub_dimension3,i If the deviation value error is within the deviation value range: The result is considered valid if the j-th sampling point of sensor s is 0 from the statistical perspective of sub-dimension 3, denoted as . Otherwise, the result is invalid, i.e. sensor i at the jth sampling point is 1 in the statistical angle of sub-dimension 3, denoted as

[0072] Timing angle: the deviation value of each time point of sensor s is judged by using a certain timing anomaly detection method such as statistical process control ewma control chart method, if the deviation value is between the upper limit and the lower limit of the jth sampling point: The result is valid, i.e. sensor s at the jth sampling point is 0 in the timing angle of sub-dimension 3, denoted as Otherwise, the result is invalid, i.e. sensor s at the jth sampling point is 1 in the timing angle of sub-dimension 3, denoted as The results of sub-dimension 3 on sensor i are processed by or operation, and the formula is as follows:

[0073]

[0074] The abnormal detection results of each sampling point of sub-dimension 3 on sensor s are output, and the abnormal detection results of each sampling point of n sensors on sub-dimension 3 are output.

[0075] In step 5, the baseline model training stage processing includes: traversing each sensor that can be regressed respectively, randomly dividing training set and test set from historical data (fault-free samples), processing independent variable parameters of training set (independent variable selection parameter set is all sensor readings related to cross section on the interface physical space of regressed sensor), dependent variable parameter is regressed sensor data, input random forest model for training, save the trained random forest model, get baseline model of each sensor on sub-dimension 4; based on historical data, the normal distribution of deviation value is established, the interval of sensor i deviation value is defined as: upper sub_dimension4,i limit upper

[0076]

[0077] Wherein represents the mean of historical deviation value of sub-dimension 4 sensor i, represents the standard deviation of sub-dimension 4 sensor i deviation value;

[0078] The lower limit lower sub_dimension4,i is:

[0079]

[0080] Statistical anomaly detection is carried out based on the deviation value distribution of each sensor model;

[0081] The data validation stage processing includes: selecting a sensor i steady-state response reading, based on all sensor readings of the relevant cross-section on the physical space of the interface where sensor i is located, using the sensor in the sub-dimension 4 baseline model to calculate the regression of the sensor reading, the deviation value of sensor i at the jth sampling point in sub-dimension 4 The calculation formula is:

[0082]

[0083] Wherein represents the regression value of sensor i at the jth sampling point in sub-dimension 4 using its own random forest baseline model, represents the reading of sensor i at the jth sampling point in the detection data;

[0084] Based on the different characteristics of the deviation value at each moment, the sensor abnormal situation is judged from the statistical and time sequence angles:

[0085] Statistical angle: input the deviation value of sensor i at each moment into the deviation value interval (lower sub_dimension4,i ,upper sub_dimension4,i ) of sensor i, if the deviation value error is within the deviation value interval range: The determination result is valid, that is, the statistical angle of sensor i at the jth sampling point in sub-dimension 4 is 0, and is recorded as Otherwise, the determination result is invalid, that is, the statistical angle of sensor i at the jth sampling point in sub-dimension 4 is 1, and is recorded as

[0086] Time sequence angle: use a certain time sequence anomaly detection method such as statistical process control ewma control chart method to judge the deviation value of sensor i at each moment, if the deviation value is between the upper limit and the lower limit of the jth sampling point: The determination result is valid, that is, the time sequence angle of sensor i at the jth sampling point in sub-dimension 4 is 0, and is recorded as Otherwise, the determination result is invalid, that is, the time sequence angle of sensor i at the jth sampling point in sub-dimension 4 is 1, and is recorded as The results of sensor i on sub-dimension 4 are processed by or operation, and the formula is as follows:

[0087]

[0088] Output the abnormal detection result of each sampling point of sensor i on sub-dimension 4;

[0089] For other sensors in the detection data, the same processing as step 5 data validation stage is performed.

[0090] In step 5, the same processing as step 5 data validation stage for other sensors under detection data is performed, specifically including: selecting a sensor s steady-state response reading, based on all sensor readings of the relevant cross section on the physical space of the interface where sensor s is located, using the sensor in sub-dimension 4 baseline model to calculate the regression of the sensor reading, the deviation value of sensor s at the jth sampling point in sub-dimension 4 The specific calculation formula is:

[0091]

[0092] Wherein, represents the regression value of sensor s at the jth sampling point in sub-dimension 4 using its own random forest baseline model, represents the reading of sensor s at the jth sampling point in the detection data;

[0093] Based on the different characteristics of the deviation value at each moment, the sensor abnormal situation is judged from the statistical and time sequence angles: statistical angle: input the deviation value of sensor s at each moment into the deviation value interval (lower sub_dimension4,i ,upper sub_dimension4,i ) of sensor s, if the deviation value error is within the deviation value interval range: The determination result is valid, that is, the statistical angle of sensor s at the jth sampling point in sub-dimension 4 is 0, denoted as Otherwise, the determination result is invalid, that is, the statistical angle of sensor s at the jth sampling point in sub-dimension 4 is 1, denoted as

[0094]

[0095] Time sequence angle: use a certain time sequence anomaly detection method such as statistical process control ewma control chart method to judge the deviation value of sensor s at each moment, if the deviation value is between the upper limit and the lower limit of the jth sampling point: The determination result is valid, that is, the time sequence angle of sensor s at the jth sampling point in sub-dimension 4 is 0, denoted as Otherwise, the determination result is invalid, that is, the time sequence angle of sensor s at the jth sampling point in sub-dimension 4 is 1, denoted as The results of sensor s on sub-dimension 4 are processed by or operation, and the formula is as follows:

[0096]

[0097] Output the abnormal detection result of each sampling point of sensor s on sub-dimension 4, and output the abnormal detection result of each sampling point of n sensors on sub-dimension 4.

[0098] In step 6, the data validation phase is performed, which specifically includes: selecting a sensor i steady-state response reading (synchronous step 2 sensor extraction sequence in sub-dimension 1), inputting the original time series data of the sensor at each time into a certain time series anomaly detection method such as statistical process control ewma control chart method for judgment, if the deviation value is between the upper limit and the lower limit of sampling point j, the determination result is valid, that is, the jth sampling point of sensor i is 0 in the time series angle of sub-dimension 5, and is recorded as Otherwise, the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the time series angle of sub-dimension 5, and is recorded as

[0099] The same processing as step 6 is performed on the detection data of other sensors, which specifically includes: selecting a sensor s steady-state response reading (synchronous step 2 sensor extraction sequence in sub-dimension 1), inputting the original time series data of the sensor at each time into a certain time series anomaly detection method such as statistical process control ewma control chart method for judgment, if the deviation value is between the upper limit and the lower limit of sampling point j, the determination result is valid, that is, the jth sampling point of sensor s is 0 in the time series angle of sub-dimension 5, and is recorded as Otherwise, the determination result is invalid, that is, the jth sampling point of sensor s is 1 in the time series angle of sub-dimension 5, and is recorded as The abnormal detection results of each sampling point of the n sensors in sub-dimension 5 are output.

[0100] Step 7 includes: using the following formula to verify whether the sensor reading is accurate:

[0101]

[0102] wherein, is a Boolean value, w=1, 2, 3, 4, 5, k here and s and i in the previous text represent the sensor number, the value range is 1- n, indicating that the kth sensor data validates the wth sub-dimension judgment result in the jth sampling point, if is 1, indicating that the kth sensor reading is abnormal, if is 0, indicating that the kth sensor reading is normal; is a Boolean value, indicating the data validation result of the kth sensor in the jth sampling point, if is 1, indicating that the data validation result of the kth sensor in the jth sampling point is inaccurate, if is 0, indicating that the data validation result of the kth sensor in the jth sampling point is accurate.

[0103] In the present application, step 3 can be used for anomaly detection for 2, 3, 4 sub-dimensions, the specific difference lies in that the selection of independent variable parameters of the training set in the baseline model training stage leads to different trained models, and the following is the analysis according to the sub-dimensions 2, 3, 4 respectively:

[0104] For sub-dimension 2: in the baseline model training stage, the independent variable parameter set is the reading of other sensors on the same section as the regression sensor and the same rake;

[0105] For sub-dimension 3: in the baseline model training stage, the independent variable parameter set is the reading of the top 10 sensors associated with the regression sensor on different sections;

[0106] For sub-dimension 4: in the baseline model training stage, the independent variable parameter set is the reading of all sensors on the relevant section in the physical space of the interface where the regression sensor is located (the following is an example):

[0107] (1) Fan section correlation analysis (to regress the reading of a sensor measuring the total temperature / total pressure at the engine inlet, the reading of a sensor measuring the total temperature / total pressure at the fan outlet can be selected as the independent variable parameter set).

[0108] (2) Compressor section correlation analysis (to regress the reading of a sensor measuring the total temperature / total pressure at the fan outlet, the reading of a sensor measuring the total temperature / total pressure at the compressor outlet can be selected as the independent variable parameter set).

[0109] (3) Combustion chamber-turbine section correlation analysis (to regress the reading of a sensor measuring the total temperature / total pressure at the compressor outlet, the reading of a sensor measuring the total temperature / total pressure at the low-pressure turbine outlet can be selected as the independent variable parameter set).

[0110] (4) Bypass parameter correlation analysis (to regress the reading of a sensor measuring the total temperature / total pressure at the bypass inlet, the reading of a sensor measuring the total temperature / total pressure at the bypass outlet can be selected as the independent variable parameter set).

[0111] In summary, the present application solves the problem that it is difficult to implement multi-sensor data validation on an aero-engine test stand, based on the time and space dimensions of engine experimental data, deeply mines the correlation between the response readings of multiple sensors, intelligently validates the effectiveness of sensor readings, and provides data support and important protection for the safety of test systems and aircraft.

[0112] Beneficial effects: the present application solves the problem that it is difficult to implement multi-sensor data validation on an aero-engine test stand, based on the time and space dimensions of engine experimental data, deeply mines the correlation between the response readings of multiple sensors, intelligently validates the effectiveness of sensor readings, and provides data support and important protection for the safety of test systems and aircraft. BRIEF DESCRIPTION OF DRAWINGS

[0113] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0114] Figure 1 This is a flowchart of the present invention.

[0115] Figure 2 This is a schematic diagram of the cross-section of the measuring point inside the engine in an embodiment of the present invention.

[0116] Figure 3 This is a schematic diagram of the distribution of the rake sensors at the fan inlet section in an embodiment of the present invention.

[0117] Figure 4 This is a schematic diagram illustrating the anomaly detection based on the timing data of a sensor under a certain test condition on an engine test bench, as shown in an embodiment of the present invention.

[0118] Figure 5 This is a schematic diagram summarizing the results of multi-sensor temporal and spatial physical dimension judgments under a certain test condition on the engine test bench in an embodiment of the present invention. Detailed Implementation

[0119] like Figure 1 , Figure 2 As shown, this invention provides a method for spatiotemporal correlation detection of multi-sensor data from an aero-engine test stand, comprising the following steps:

[0120] Step 1: Install n sensors inside the aero-engine to identify data from the aero-engine under different test conditions on the test bench (such as idle, N2: 90%, N2: 95%, high power, etc.). Extract the response readings of each time-series sampling point of the n sensors under the steady-state state of one of the conditions (such as high power state). (If there are a total of m sampling points, the extracted data matrix shape is m*n, that is, the response reading of each sensor at each time-series sampling point).

[0121] Step 2, baseline model training phase for sub-dimension 1: put historical data into the box plot model to measure the upper and lower limits, and sequentially process the difference of data of the same section, radius and sampling point. Replace each difference value with the original data of the sampling point, and do the same for other sampling points and sections. Finally, input all sequential difference values of data of the same section, sampling point and radius into the box plot model, which is called baseline model 1, to obtain the upper and lower limits of different sections based on historical data. Data validation phase: select a stable sensor response reading (generally in the order of engine section relationship, if the same section is measured by the rake number, take the sensor from the flow field center to the outside), and still need to sequentially process the difference of data of the test sample, section, radius and sampling point. According to the model (upper and lower limits of different sections) corresponding to the sensor in baseline model 1, judge whether the sensor is abnormal; when other sensors are confirmed, do the same processing as step 2 in the data validation phase for other sensors under the same working condition, and then output the abnormal detection results of each sampling point of n sensors in sub-dimension 1;

[0122] Step 3, baseline model training phase for sub-dimensions 2, 3 and 4: traverse each regressionable sensor respectively, randomly divide the training set and test set from all high-power state data (no fault sample), process the independent variable parameters (different sub-dimensions and different independent variable parameters) of the training set, and input the regression sensor data into the random forest model for training. Save the trained model, so the baseline model of each sensor in this sub-dimension can be obtained. Based on the normal distribution of the deviation value of the historical normal sample data, the interval of the sensor deviation value is defined as the upper limit of the mean value + 3.5*standard deviation of the sensor historical deviation value, and the lower limit of the mean value - 3.5*standard deviation of the sensor historical deviation value, and then the statistical abnormal detection can be carried out based on the deviation value distribution of each sensor model. Data validation phase: select a stable sensor response reading, based on the data of the same section and rake of the sensor, use the sensor baseline model in this sub-dimension to regress the sensor reading, and judge the sensor abnormality from the statistical and time sequence angles based on the different characteristics of each time deviation value: statistical angle: input the deviation value of the sensor at each time into the model, if the value error is within the deviation value interval, the result is effective (0), otherwise it is invalid (1). Time sequence angle: input the deviation value of the sensor at each time into the model, and use a time sequence anomaly detection method such as statistical process control ewma control chart method to judge, if the deviation value is between the upper limit (ucl) and the lower limit (lcl) of the point, the data of the time point is effective (0), otherwise it is invalid or abnormal (1). Do the same processing as step 3 for other sensors under the same working condition, and output the abnormal detection results of each sampling point of n sensors in this sub-dimension;

[0123] Step 4, data validation stage for sub-dimension 5: select a sensor steady-state response reading (sensor extraction sequence of sub-dimension 1 in step 2), input the original data of each time point of the sensor into a certain time series anomaly detection method such as statistical process control ewma control chart method for judgment, if the deviation value is between the upper limit (ucl) and the lower limit (lcl) of the point, it is considered that the data of the time point is valid (0), otherwise it is recorded as invalid or abnormal (1); the same processing as step 4 data validation stage is carried out for other sensors under the same working condition, and the abnormal detection results of n sensors in sub-dimension 5 are output;

[0124] Step 5, combine sub-dimension 1, 2, 3 results into spatial dimension results, and sub-dimension 4, 5 abnormal detection results as physical and time dimension results, combine abnormal detection results from spatial, physical and time three large dimensions, and validate sensor data, and perform dimension analysis on abnormal sensor readings, and reconstruct data according to abnormal dimension.

[0125] Further, the engine section in step 1 includes: engine inlet section, fan outlet section, compressor outlet section, low pressure turbine outlet section, bypass inlet section, and bypass outlet section as shown in Figure 2 (1-6) are respectively the physical space positions of the above-mentioned sections in the engine, and different numbers of sensors are arranged in different ways on each section to measure the total pressure and total temperature of the section; the abnormal detection of sub-dimension 1-4 is based on the engine section.

[0126] Further, step 2 includes:

[0127] The upper and lower limit thresholds are trained based on the historical data of the same week sensor of a section using the box plot model. For the same section and radius dimension baseline model, select a sensor and extract all response readings in this period. Input the extracted time series data into a suitable baseline model to validate the sensor reading of this test run. In this way, the box plot model error caused by insufficient data is avoided.

[0128] In the model training stage, for example, at the fan outlet section (PtM), there are 5 measuring rakes at this section, and each measuring rake has 3 measuring points. That is, at one sampling point of this section, there are only 3 data that can be used for abnormality judgment by the box plot. The data sample is small. Therefore, the historical data needs to be put into the box plot model to measure the upper and lower limits. However, the values of the same sensor under different external conditions (such as altitude, temperature, etc.) can be very different. In order to eliminate the influence of different external conditions, the data of the same section, the same radius and the same sampling point need to be sequentially processed. Still taking the fan outlet section (PtM) as an example, the specific processing method is that for the same sampling point, the values of PtM_01_01-PtM_01_02, PtM_01_02-PtM_01_03 and PtM_01_03-PtM_01_01 are calculated (the value of PtM_01_01 is the reading of the first measuring point (the sensor closest to the flow field center) of the first measuring rake at the fan outlet section, and similarly, the value of PtM_01_02 is the reading of the second measuring point sensor of the first measuring rake at the fan outlet section). Replace the original three data of the sampling point with these three difference values, and do the same for other sampling points and sections. Finally, all the sequentially processed data of the same sampling point, the same section and the same radius are input into the box plot model to obtain the upper and lower limits of the section based on the historical data.

[0129] In the data validation stage, the values of PtM_01_01-PtM_01_02, PtM_01_02-PtM_01_03 and PtM_01_03-PtM_01_01 are calculated for the same sampling point, which are input into the model trained in the previous step to judge the abnormality of the three sensors at the sampling point. Assuming that PtM_01_01-PtM_01_02 and PtM_01_02-PtM_01_03 exceed the upper and lower limits of the model, then the sensor PtM_01_02 is considered to be abnormal at this sampling point in this dimension. Then, the same method is used for all sampling points in this test run, and the abnormality of all sensors at each sampling point in the PtM_01_xx section in this sub-dimension can be obtained. The same applies to the sensors on other radii of this section and the sensors on the same radius of other sections. Finally, the abnormality detection results of each sampling point of all sensors in the sub-dimension 1 are output.

[0130] Further, step 3 includes,

[0131] Taking the fan outlet interface as an example, assuming that there are 5 measuring rakes at the fan outlet section, and each measuring rake has 3 sensors. Based on the data of different radii of the same measuring rake, a suitable baseline model is selected to calculate the deviation value of the selected data reading in this step. The features of the deviation value in statistics and time sequence are extracted, and the reading of each sampling point of the sensor is verified based on this.

[0132] Still taking the fan outlet section total pressure 01_01 ('PtM_01_01') parameter as an example:

[0133] Model training stage: input parameters: input parameters of the same section and the same measuring probe as the fan outlet section total pressure 01_01 ('PtM_01_02', 'PtM_01_03', 'PtM_01_04', 'PtM_01_05' parameters) Training matrix D: (assuming that there are a total of 509 sampling points in the historical data, PtM_01_02 509 represents the sensor reading of the second measuring probe on the first measuring probe at the 509th sampling point)

[0134]

[0135] Each table represents a monitoring parameter, and each row represents a set of parameters selected from the same section and the same radius as the regression parameter in a steady state.

[0136] Test matrix:

[0137]

[0138] Output parameter: output a performance parameter prediction value (such as PtM_01_01 (i.e. fan outlet section total pressure 01_01))

[0139]

[0140] Data preprocessing: construct the independent variable training set x_train and the regression training set y_train; the independent variable test set x_test and the regression test set y_test. Model training: train the processed x_train and y_train, and save the trained model. Model testing: input x_test and y_test, and observe the output results.

[0141] Finally, according to the above method, for the data analyzed this time, the correlation model of the same measuring probe parameters in the section where all sensors are located is constructed. Based on the correlation model of the same section and the same measuring probe parameters, the deviation between the measured value of each sensor signal and the predicted value of the correlation model (i.e. the predicted value based on the same section and the same measuring probe other measuring points) can be obtained. First, the normal distribution of the deviation value is established based on the historical normal sample data, and then statistical anomaly detection can be performed based on the deviation value distribution. The specification of the deviation value interval of a certain sensor in this report is that the upper limit is the mean value of the historical deviation value of the sensor + 3.5*standard deviation of the deviation value, and the lower limit is the mean value of the historical deviation value of the sensor - 3.5*standard deviation of the deviation value.

[0142] Data validation stage:

[0143] Statistical angle: we will test data (['PtM_01_01'], ['PtM_01_02', 'PtM_01_03', 'PtM_01_04' and 'PtM_01_05']) into the model corresponding to the above sub-dimension 2, get the test data each sampling point deviation value, if the prediction result error is in the deviation value interval range, it is considered that the result is valid, otherwise it is invalid. Sampling points exceeding the threshold are recorded as abnormal situation 1 in this dimension, otherwise 0.

[0144] Timing angle: test data ('PtM_01_01', 'PtM_01_02', 'PtM_01_03', 'PtM_01_04' and 'PtM_01_05') are input into the model corresponding to the above sub-dimension 2, and the deviation value of each sampling point of the test data is obtained. A certain time series anomaly detection method such as statistical process control ewma control chart method is used for judgment (its definition formula is:

[0145] z i = λx i +(1-λ)z i-1

[0146] Where the constant λ is in the range of 0≤λ≤1, z i is the ewma statistic, that is, the weighted average of all previous sample means, and the initial value z0 of the formula is the target value μ0 of the process (that is, z0=μ0). Since ewma is not sensitive to the normality assumption of data, it is very ideal for single observation value, and in this example x i is the deviation value of each sampling point of the test data. Obviously, x i is an independent random variable with variance σ 2 , so the variance of z i is:

[0147]

[0148] ), which will generate upper limit (ucl) and lower limit (lcl) for each sampling point according to the input data. The specific formula is,

[0149]

[0150]

[0151] When i increases, lcl and ucl will stabilize to the following two values:

[0152]

[0153] If the deviation value is between the upper limit (ucl) and the lower limit (lcl) of the point, it is considered that the data of the sampling point is valid (0), otherwise it is recorded as invalid or abnormal (1).

[0154] For the sub-dimension 2 statistical angle result and the timing angle result, an 'or' operation is performed, and the abnormal detection result of each sampling point of the test data on the sub-dimension 2 is obtained.

[0155] For sub-dimension 3 and sub-dimension 4, their processing processes are similar to those of sub-dimension 2. The difference lies in that the difference in the selection of independent variable parameters of the training set in the baseline model training stage leads to different trained models. For sub-dimension 3: in the baseline model training stage, the independent variable selection parameter set is the top 10 sensor readings of the correlation on different sections of the regression sensor; for sub-dimension 4: in the baseline model training stage, the independent variable selection parameter set is all sensor readings on the relevant section in the physical space of the interface of the regression sensor; the process of the data validation stage remains unchanged.

[0156] Further, step 4 includes,

[0157] For sub-dimension 5, in the data validation stage, a sensor steady-state response reading is selected (in the sensor extraction sequence of sub-dimension 1 in step 2), and the original data of the sensor at each time is input into a certain timing abnormality detection method such as a statistical process control ewma control chart method for judgment. Unlike the timing angle of sub-dimensions 2, 3, and 4, the input of sub-dimensions 2, 3, and 4 is a deviation value, while the input selected in step 4 is original data. If the deviation value is between the upper limit (ucl) and the lower limit (lcl) of the point calculated by the ewma algorithm, it is considered that the data of the time point is valid (0), otherwise it is recorded as invalid or abnormal (1). The same processing as the data validation stage of step 4 is performed on other sensors under the working condition, and the abnormal detection results of n sensors on sub-dimension 5 at each sampling point are output.

[0158] Further, step 5 includes,

[0159] The abnormality detection results of sub-dimension 1, (sub-dimension 2, sub-dimension 3, sub-dimension 4), and sub-dimension 5 are obtained respectively by summarizing steps 2, 3, and 4. The sub-dimensions 1, 2, and 3 are combined by an OR operation to obtain the abnormality detection result of the data space dimension in this test run. The sub-dimensions 4 and 5 are output as the abnormality detection results of the physical dimension and the time dimension, respectively. In this way, the sensor reading validation can be analyzed by combining the time and space dimensions and using the OR operation to obtain the comprehensive abnormality detection result of each sensor at each sampling point (i.e., each sensor comprehensive result is a Boolean array with a length of the number of sampling points). Finally, the mean value operation is performed on each sensor comprehensive abnormality result to obtain the data score of the sensor in this test run, and the abnormality possibility of the sensor is determined according to the score. If no abnormality is found in the sensor, the sensor data is validated as effective. If the sensor has an abnormality, the invalid dimension is traced back, and the response reading of the sensor is reconstructed based on this.

[0160] Embodiment

[0161] Based on Figure 3 The engine internal measurement points are briefly described as follows:

[0162] Figure 3 is a schematic diagram of the sensor cross section set in the engine. In the diagram, the inlet 1 (engine inlet total temperature and total pressure) is before the fan, and then the fan outlet 2 (fan outlet total temperature and total pressure) is. The fan outlet is divided into two paths: the first path is the inner core, including the flow through the compressor 3 (compressor outlet total temperature and total pressure), the combustion chamber, and the high and low pressure turbine 4 (low pressure turbine outlet total temperature and total pressure); the other path enters the outer core 5 (outer core inlet total temperature and total pressure), and the gas flow at the outer core outlet 6 is mixed with the gas flow at the low pressure turbine outlet in a mixer.

[0163] After a brief introduction to the engine internal sensor setting cross section, the distribution method of the sensors on each cross section is introduced. The measurement rake layout and the sensor distribution method are different on each cross section. (Taking the fan outlet cross section as an example), it is assumed that there are 4 measurement rakes on the fan outlet cross section, and 3 sensors are distributed on each measurement rake (as shown in Figure 3 Each sensor can measure the temperature and pressure of the flow field at the position.

[0164] Taking step 3 of the application as an example, a training set and a test set are randomly divided from historical data (non-fault samples), the independent variable parameters of the training set are processed (the sub-dimension 2 independent variable selection parameter set is the reading of other sensors on the same section as the regression sensor), the dependent variable parameter is the regression sensor data, a random forest model is input for training, the trained random forest model is saved, and a baseline model of each sensor on sub-dimension 2 is obtained; a normal distribution of bias values is established based on historical data, and the bias value interval of sensor i is defined as: upper limit upper sub_dimension,i is:

[0165]

[0166] wherein represents the mean of the historical bias value of the sub-dimension 2 sensor i, represents the standard deviation of the bias value of the sub-dimension 2 sensor i;

[0167] lower limit lower sub_dimension2,i is:

[0168]

[0169] Statistical anomaly detection is performed based on the bias value distribution of each sensor model;

[0170] The data validation stage processing includes: selecting a steady-state response reading of a sensor i, based on the correlation of the top 10 sensor data on different sections where the sensor is located, using the sub-dimension 2 baseline model of the sensor to regress and calculate the sensor reading, and the bias value of the sensor i on the sub-dimension 2 at the jth sampling point is The calculation formula is:

[0171]

[0172] wherein represents the regression value of the sensor i on the sub-dimension 2 at the jth sampling point using its own random forest baseline model, represents the reading of the sensor i at the jth sampling point in the detection data;

[0173] Time sequence angle: the bias value of the sensor i at each time is judged using the existing ewma model, if the bias value is between the upper limit and the lower limit of the jth sampling point: the determination result is valid, that is, the jth sampling point of the sensor i on the sub-dimension 2 is 0 in the time sequence angle, and is recorded as otherwise, the determination result is invalid, that is, the jth sampling point of the sensor i on the sub-dimension 2 is 1 in the time sequence angle, and is recorded as (in Figure 4The middle blue dot is The upper red dot line is composed of The lower red dot line is composed of Similarly, we can see that Figure 4 When j = 32, 33, the deviation value exceeds the upper limit, so we can get the abnormal detection result of the timing angle of the sensor i in the sub-dimension 2 Wherein The others are 0

[0174] In the specific operation, according to different dimensions, the suitable baseline model and the appropriate input data input model are determined. In this way, for the same sensor, five corresponding abnormal detection schemes can be obtained through steps 2, 3 and 4. Generalized to all measuring points of the engine, that is, for each measuring point, five corresponding abnormal detection schemes can be obtained through the five sub-dimensions, (integrated and merged into three large dimensions (space, physics, time)). Assuming that the engine has n measuring points in total, the results of all the results are combined to obtain a 3*n-dimensional matrix (as shown in Figure 5 The column of the matrix is operated or, and the abnormal condition of each measuring point of the engine at each sampling point can be determined. Due to the instability of the flow field and other factors, it cannot be determined that the sensor has a problem in this test run, so the average value of all sampling points is used to represent the abnormal possibility of the sensor. Assuming that the engine test bed has 80 measuring points in total for one test run, the possibility of abnormality of the No. 1 sensor in the space dimension (sub-dimension 1) is very large, so the response reading of the sensor is proved to be invalid. The data reconstruction of the sensor needs to be performed according to the output of the space dimension (sub-dimension 1).

[0175] If the sensor is abnormal, the dimension where the abnormality occurs is traced back, and the data reconstruction is performed according to the abnormal dimension.

[0176] In the specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and the computer program can run the invention content of the aviation engine test bed multi-sensor data space-time correlation detection method provided by the present application and part or all steps in each embodiment when executed by the data processing unit. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM) and the like.

[0177] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present application can be implemented by means of a computer program and a corresponding general hardware platform. Based on such understanding, the technical solutions in the embodiments of the present application can be embodied in the form of a computer program, i.e., a software product, which can be stored in a storage medium, including a plurality of instructions for causing a device (which can be a personal computer, a server, a single-chip microcomputer, a MUU, or a network device) including a data processing unit to execute the method described in various embodiments or some parts of the embodiments of the present application.

[0178] The present application provides an aero-engine test stand multi-sensor data space-time correlation detection method. There are many methods and ways to achieve the technical solutions of the present application. The above description is only the preferred embodiment of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application. The components not explicitly described in the embodiments can be implemented by using existing technology.

Claims

1. A method for detecting spatio-temporal correlation of multi-sensor data of an aero-engine test stand, characterized in that, Comprise the following steps: Step 1, n sensors are arranged in the aero-engine, the data of the aero-engine in different tests inside the test bench are identified, the response reading of each time sequence sampling point of n sensors in a steady state under one working condition is extracted as historical data for use in the baseline model training stage; one steady state data of a new test under any working condition is used as detection data for use in the data validation stage, and the sensor data is divided into three large dimensions of space, physics and time, based on the three large dimensions, it is divided into five sub-dimensions of same cross-section, same radius and different harrows, same cross-section, same harrow and different radius, different cross-section, parameter correlation and original time sequence, which are respectively recorded as sub-dimensions 1, 2, 3, 4 and 5, wherein the sub-dimensions 1, 2 and 3 belong to the space dimension, the sub-dimension 4 belongs to the physics dimension, and the sub-dimension 5 belongs to the time dimension; Step 2, for sub-dimension 1, baseline model training stage processing and data validation stage processing are carried out; Step 3, for sub-dimension 2, baseline model training stage processing and data validation stage processing are carried out; Step 4, for sub-dimension 3, baseline model training stage processing and data validation stage processing are carried out; Step 5, for sub-dimension 4, baseline model training stage processing and data validation stage processing are carried out; Step 6, for sub-dimension 5, data validation stage processing is carried out; Step 7, the results of sub-dimensions 1, 2 and 3 are combined as the space dimension result, the abnormal detection results of sub-dimensions 4 and 5 are respectively taken as the physics and time dimension results, the abnormal detection results are combined from the three large dimensions of space, physics and time, and the sensor data is validated, and the abnormal sensor reading is analyzed in dimension, and the data is reconstructed according to the abnormal dimension.

2. The method of claim 1, wherein, In step 2, the baseline model training stage processing includes: the following processing is performed on all sampling points and cross sections of the historical data: sequentially difference processing is performed on the data of the same cross section, the same radius and the same sampling point; each difference value replaces each data of the original sampling point, and finally all the sequentially difference processed data of the same sampling point, the same cross section and the same radius are input into a box plot, which is called baseline model 1, to obtain the upper and lower limits of the cross section based on the historical data; The data validation stage processing includes: selecting a steady state response reading of a sensor A in the detection data, sequentially difference processing is performed on the data of the same cross section, the same radius and the same sampling point, and whether the sensor is abnormal is judged according to the model corresponding to the sensor A in the baseline model 1; in the validation of other sensors, the same processing as step 2 is performed on the other sensors in the detection data, and then the abnormal detection results of each sampling point of n sensors in sub-dimension 1 are output.

3. The method of claim 2, wherein, In step 3, the baseline model training stage processing includes: traversing each sensor capable of regression respectively, randomly dividing training set and test set from historical data, processing independent variable parameters of the training set, dependent variable parameters being regression sensor data, inputting random forest model for training, saving the trained random forest model, and obtaining random forest baseline model of each sensor in sub-dimension 2; establishing normal distribution of bias value based on historical data, the bias value interval of sensor i being defined as: upper limit upper sub_dimension2,i is: wherein denotes the standard deviation of the sub-dimension 2 sensor i bias values; lower limit sub_dimension2,i is: Statistical abnormal detection is carried out based on the model deviation value distribution of each sensor; The data validation stage processing includes: selecting a sensor i steady-state response reading, based on the data of the cross section where the sensor is located and the different radii of the measuring rake, using the sensor in the sub-dimension 2 baseline model to regress the sensor reading, and the deviation value of the sensor i at the jth sampling point in the sub-dimension 2 The calculation formula is: wherein represents the regression value of sensor i at the jth sampling point on the 2nd sub-dimension using its own random forest baseline model, represents the reading of sensor i at the jth sampling point in the detection data; The abnormal condition of the sensor is judged from the statistical and time sequence angles based on different characteristics of each time deviation value: Statistical angle: input the deviation value of sensor i at each time into the deviation value interval (lower sub_dimension2,i ,upper sub_dimension2,i ) of sensor i, if the deviation value error is within the deviation value interval range: then the determination result is valid, that is, the jth sampling point of sensor i is 0 in the statistical angle of sub-dimension 2, denoted as otherwise the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the statistical angle of sub-dimension 2, denoted as Time sequence angle: the deviation value of sensor i at each time is judged by using the statistical process control ewma control chart method, and the formula of the statistical process control ewma control chart method is as follows: z t = λx t + (1 - λ)z t-1 Where the constant λ takes values ​​in the range of 0 ≤ λ ≤ 1, z t It is the Ewma statistic, x t Let z0 be the deviation value of the test data sample point t of sensor i, where t ranges from 0 to m, m is the number of sampling points in one experiment, and t is an integer. The initial value of the Ewma statistic z0 is taken as the mean value μ0 of the test data of sensor i; t Let each be an independent random variable with variance σ. 2 , z t variance for: The process control ewma control chart method generates upper limit ucl and lower limit lcl of each sampling point according to the input data, and the formula is as follows: where L is a constant, σ is an independent random variable x t of the standard deviation; as i increases, the lcl and ucl will stabilize to the following two values: If the deviation value is between the upper limit and the lower limit : then the determination result is valid, that is, the jth sampling point of sensor i is 0 in the time sequence angle of sub-dimension 2, denoted as Otherwise, the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the time sequence angle of sub-dimension 2, denoted as The or operation processing is performed on the results of sensor i on sub-dimension 2, and the formula is as follows: Output each sampling point of the sensor i on the sub-dimension 2 exception detection result; For the detection data under other sensors do the same processing in step 3.

4. The method of claim 3, wherein, In step 3, the same processing of other sensors in step 3 is carried out, which specifically includes: selecting a sensor s steady-state response reading, based on the data of the same section and different radii of the rake of the sensor, using the sensor in the sub-dimension 2 baseline model to regress the sensor reading, and the specific calculation formula of the deviation value of the sensor s on the sub-dimension 2 jth sampling point is: Based on the different characteristics of the deviation value at each moment, the sensor abnormality is judged from the statistical and time sequence angles: Statistical angle: input the deviation value of sensor s at each moment into the deviation value interval (lower sub_dimension2,i ,upper sub_dimension2,i ) of sensor s, if the error of the deviation value is within the range of the deviation value interval: then the determination result is valid, that is, the jth sampling point of sensor s is 0 in the statistical angle of sub-dimension 2, denoted as otherwise the determination result is invalid, that is, the jth sampling point of sensor s is 1 in the statistical angle of sub-dimension 2, denoted as Timing angle: the deviation value of each time of sensor s is judged by statistical process control ewma control chart method, if the deviation value is between the upper limit and the lower limit of sampling point j: and then the determination result is valid, that is, the jth sampling point of sensor s is 0 in the timing angle of sub-dimension 2, denoted as otherwise, the determination result is valid, that is, the jth sampling point of sensor s is 1 in the timing angle of sub-dimension 2, denoted as The results of the sensors s on the sub-dimension 2 are processed by or operation, and the formula is as follows:​ Output each sampling point of the sensor i on the sub-dimension 2 exception detection result; finally output each sampling point of the n sensors on the sub-dimension 2 exception detection result.

5. The method of claim 4, wherein, In step 4, the baseline model training stage processing includes: traversing each sensor capable of regression respectively, randomly dividing training set and test set from historical data, processing independent variable parameters of the training set, dependent variable parameters being regression sensor data, inputting random forest model for training, saving the trained random forest model, and obtaining the baseline model of each sensor in sub-dimension 3; establishing normal distribution of bias value based on historical data, the bias value interval of sensor i being defined as: upper limit upper sub_dimension3,i is: wherein denotes the mean value of the historical bias values of the sub-dimension 3 sensor i, denotes the standard deviation of the bias values of the sub-dimension 3 sensor i; lower limit sub_dimension3,i is: Statistical anomaly detection based on the deviation value distribution of each sensor model; The data validation stage processing includes: selecting a sensor i steady-state response reading, using the sensor in the sub-dimension baseline model to regress the sensor reading, and calculating the deviation value of sensor i at the jth sampling point in the sub-dimension 3 The calculation formula is: wherein represents the regression value of sensor i at the jth sampling point on the 3rd sub-dimension using its own random forest baseline model, represents the reading of sensor i at the jth sampling point in the detection data; Based on the different characteristics of the deviation value at each moment, the sensor abnormality is judged from the statistical and time sequence angles: Statistical angle: input the deviation value of sensor i at each time into the deviation value interval (lower sub_dimension3,i ,upper sub_dimension3,i ) of sensor i, if the error of the deviation value is within the range of the deviation value interval , the determination result is valid, that is, the jth sampling point of sensor i is 0 in the statistical angle of sub-dimension 2, denoted as , otherwise the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the statistical angle of sub-dimension 2, denoted as Timing angle: the deviation value of each time of sensor i is judged by statistical process control ewma control chart method, if the deviation value is between the upper limit and the lower limit of sampling point j: and then the determination result is valid, that is, the jth sampling point of sensor i is 0 in the timing angle of sub-dimension 3, recorded as otherwise, the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the timing angle of sub-dimension 3, recorded as ​ Or operation processing is carried out on the sensor i result on the sub-dimension 3, and the formula is as follows: Output each sampling point of the sensor i on the sub-dimension 3 exception detection result; For the detection data under other sensors do the same processing in step 4 data confirmation stage.

6. The method of claim 5, wherein, In step 4, the same data validation process is applied to the other sensors for the test data, including: selecting a steady-state response reading of a sensor s, using the sensor's baseline model in sub-dimension 3 to regress the sensor reading, and calculating the bias value of the sensor s at the jth sampling point in sub-dimension 3 The calculation formula is as follows: Based on the different characteristics of the deviation value at each moment, the sensor abnormal situation is judged from the statistical and timing angles: statistical angle: the deviation value of sensor s at each moment is input into the deviation value interval (lower sub_dimension3,i ,upper sub_dimension3,i ) of sensor s, if the deviation value error is within the deviation value interval range: The determination result is valid, that is, the jth sampling point of sensor s is 0 in the statistical angle of sub-dimension 3, and is recorded as Otherwise, the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the statistical angle of sub-dimension 3, and is recorded as Timing angle: the deviation value of each time of sensor s is judged by statistical process control ewma control chart method, if the deviation value is between the upper limit and the lower limit of sampling point j: and then the determination result is valid, that is, the jth sampling point of sensor s is 0 in the timing angle of sub-dimension 3, denoted as otherwise, the determination result is invalid, that is, the jth sampling point of sensor s is 1 in the timing angle of sub-dimension 3, denoted as The results of sub-dimension 3 on sensor i are processed by or operation, and the formula is as follows:​ Output each sampling point of the sensor s on the sub-dimension 3 exception detection result, and output each sampling point of the n sensors on the sub-dimension 3 exception detection result.

7. The method of claim 6, wherein, In step 5, the baseline model training stage processing includes: traversing each sensor capable of regression respectively, randomly dividing training set and test set from historical data, processing independent variable parameters of the training set, dependent variable parameters being regression sensor data, inputting random forest model for training, saving the trained random forest model, and obtaining the baseline model of each sensor in sub-dimension 4; establishing normal distribution of bias value based on historical data, the bias value interval of sensor i being defined as: upper limit upper sub_dimension4,i is: wherein represents the mean value of the historical bias values of the sub-dimension 4 sensor i, represents the standard deviation of the bias values of the sub-dimension 4 sensor i; lower limit sub_dimension4,i is: Statistical anomaly detection based on the deviation value distribution of each sensor model; The data validation stage processing includes: selecting a sensor i steady-state response reading, based on all sensor readings of the relevant cross section on the physical space of the interface where sensor i is located, using the sensor in the sub-dimension 4 baseline model to calculate the regression of the sensor readings, and the deviation value of sensor i at the jth sampling point in the sub-dimension 4 The calculation formula is: wherein represents the regression value of sensor i at the jth sampling point on the 4th sub-dimension using its own random forest baseline model, represents the reading of sensor i at the jth sampling point in the detection data; Based on the different characteristics of the deviation value at each moment, the sensor abnormality is judged from the statistical and time sequence angles: Statistical angle: input the deviation value of sensor i at each time into the deviation value interval (lower sub_dimension4,i ,upper sub_dimension4,i ) of sensor i, if the deviation value error is within the deviation value interval range: then the determination result is valid, that is, the jth sampling point of sensor i is 0 in the statistical angle of sub-dimension 4, denoted as otherwise, the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the statistical angle of sub-dimension 4, denoted as Timing angle: the deviation value of each time of sensor i is judged by statistical process control ewma control chart method, if the deviation value is between the upper limit and the lower limit of sampling point j: and then the determination result is valid, that is, the jth sampling point of sensor i is 0 in the timing angle of sub-dimension 4, denoted as otherwise, the determination result is invalid, that is, the jth sampling point of sensor i is 1 in the timing angle of sub-dimension 4, denoted as The results of sub-dimension 4 on sensor i are processed by or operation, and the formula is as follows:​ Output each sampling point of the sensor i on the sub-dimension 4 exception detection result; For the detection data under other sensors do the same processing in step 5 data confirmation stage.

8. The method of claim 7, wherein, In step 5, the same process is applied to the other sensors under the detected data, specifically including: selecting a sensor s steady-state response reading, based on all sensor readings on the relevant cross section of the physical space of the interface where sensor s is located, using the sensor in the sub-dimension 4 baseline model to calculate the sensor readings, the deviation value of sensor s at the jth sampling point in the sub-dimension 4 The specific calculation formula is as follows: Based on the different characteristics of the deviation value at each moment, the sensor abnormal situation is judged from the statistical and timing angles: statistical angle: the deviation value of sensor s at each moment is input into the deviation value interval (lower sub_dimension4,i ,upper sub_dimension4,i ) of sensor s, if the deviation value error is within the deviation value interval range: The judgment result is valid, that is, the jth sampling point of sensor s is 0 in the statistical angle of sub-dimension 4, which is recorded as Otherwise, the judgment result is invalid, that is, the jth sampling point of sensor s is 1 in the statistical angle of sub-dimension 4, which is recorded as Timing angle: the deviation value of each time of sensor s is judged by statistical process control ewma control chart method, if the deviation value is between the upper limit and the lower limit of sampling point j: and then the determination result is valid, that is, the jth sampling point of sensor s is 0 in the timing angle of sub-dimension 4, denoted as otherwise, the determination result is invalid, that is, the jth sampling point of sensor s is 1 in the timing angle of sub-dimension 4, denoted as The results of the sensors s on the sub-dimension 4 are processed by or operation, and the formula is as follows:​ Output each sampling point of the sensor s on the sub-dimension 4 exception detection result, and output each sampling point of the n sensors on the sub-dimension 4 exception detection result.

9. The method of claim 8, wherein, In step 6, the data validation phase is performed, specifically including: selecting a sensor i steady-state response reading, inputting the original time series data of each time of the sensor into the statistical process control ewma control chart method for judgment, if the deviation value is between the upper limit and the lower limit of the sampling point j , the determination result is valid, that is, the jth sampling point of the sensor i is 0 in the time series angle of the sub-dimension 5, and is recorded as Otherwise, the determination result is invalid, that is, the jth sampling point of the sensor i is 1 in the time series angle of the sub-dimension 5, and is recorded as ​ For the other sensors under the detection data, the same processing as step 6 data validation stage is carried out, which specifically includes: selecting a sensor s steady-state response reading, inputting the original time series data of each time of the sensor into the statistical process control ewma control chart method for judgment, if the deviation value is between the upper limit and the lower limit of the sampling point j , the determination result is valid, that is, the jth sampling point of the sensor s is 0 in the time sequence angle of the sub-dimension 5, and is recorded as Otherwise, the determination result is invalid, that is, the jth sampling point of the sensor s is 1 in the time sequence angle of the sub-dimension 5, and is recorded as Output the abnormal detection result of each sampling point of the n sensors in the sub-dimension 5.​ 10. The method of claim 9, wherein, Step 7 includes: using the following formula to confirm whether the sensor reading is accurate: wherein, is a Boolean value, w = 1, 2, 3, 4, 5, indicating that the kth sensor data confirms the wth sub-dimension judgment result at the jth sampling point, if is 1 indicating that the kth sensor reading is abnormal, if is 0 indicating that the kth sensor reading is normal; is a Boolean value, indicating that the kth sensor data confirms the result at the jth sampling point, if is 1 indicating that the kth sensor data confirms the result at the jth sampling point is inaccurate, if is 0 indicating that the kth sensor data confirms the result at the jth sampling point is accurate.

Images