Equipment service process-oriented abnormal sensing information online identification and self-calibration method
By using a single-step prediction model and graph convolution module to identify and self-calibrate abnormal sensing information during the service process of complex equipment, the identification and calibration problems in the prior art are solved, and the stability and sensing information of the service process of complex equipment are achieved. Quality assurance.
Patent Information
- Application Number
- CN202410940899.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-12
AI Technical Summary
The prior art is difficult to quickly and efficiently identify and self-calibrate abnormal sensing information in the industrial Internet of Things, resulting in the impact of the stability and quality of sensing information monitoring of complex equipment service processes.
A single-step prediction model is used to determine the abnormal situation through the sensor's historical time sensing information sampling value and predicted value, and use the predicted value instead of the sampling value for self-calibration during the abnormality period, and combine the graph learning module and graph convolution module for abnormal type identification. .
It realizes online accurate identification and self-calibration of abnormal sensing information in the service process of complex equipment, ensures monitoring stability and sensor information quality, and can promptly detect and handle abnormal situations.
Smart Images

Figure CN118936540A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of online recognition and self-calibration of abnormal sensor information, and in particular to an online recognition and self-calibration method of abnormal sensor information oriented to the service process of equipment. Background Art
[0002] In recent years, with the rapid development of the Industrial Internet of Things, a large number of sensing devices (also called sensors) have been widely used in the service process of complex equipment. These sensing devices monitor the service performance of complex equipment by real-time monitoring of various parameters in the service process of complex equipment, such as temperature, pressure, flow, etc., to achieve precise control and management of complex equipment, and provide important sensor information support for subsequent in-depth mining and analysis of the service process of complex equipment. However, since the sensor equipment is extremely susceptible to interference such as harsh service environment and human damage, the collected sensor information is unreliable. Therefore, it is of great significance to quickly and efficiently identify abnormal sensor information and accurately calibrate abnormal sensor information to ensure the stability of monitoring the service process of complex equipment and improve the quality of sensor information.
[0003] At present, the research methods of sensor information anomaly monitoring at home and abroad can be roughly divided into two categories: offline monitoring methods and online monitoring methods. Offline monitoring methods mainly analyze and mine historical sensor information, identify abnormal sensor information, and perform processing such as elimination, interpolation, estimation and correction to make it closer to the actual service process. However, this type of method has problems such as high consumption of sensor information storage resources, difficulty in abnormal intervention, and low timeliness of sensor information. It is difficult to meet the real-time monitoring needs of abnormal sensor information under the background of rapid development of industrial Internet of Things. The online monitoring method is to compare the current sensor information with the estimated threshold through rule knowledge, statistical analysis, machine learning and other methods to achieve online monitoring of abnormal sensor information. However, this type of method lacks the ability of sensor information self-calibration under abnormal sensor information, and does not have the ability to protect the quality of sensor information from interference during abnormal periods. Summary of the invention
[0004] The purpose of the present invention is to provide an online identification and self-calibration method for abnormal sensor information in the service process of equipment, which can realize online accurate identification and self-calibration of abnormal sensor information, and ensure the stability of monitoring the service process of complex equipment and the quality of sensor information.
[0005] To achieve the above object, the present invention provides the following solutions:
[0006] An online recognition and self-calibration method for abnormal sensor information during equipment service is provided. For each sensor arranged on complex equipment, the following steps are performed:
[0007] The sensor information sampling values of the sensor at multiple consecutive historical moments before time t are selected to form a prediction sequence of the sensor at time t; the sensor is used to monitor the service performance of complex equipment; the last moment of the multiple consecutive historical moments is time t-1;
[0008] Taking the prediction sequence of the sensor at time t as input, using a single-step prediction model to determine the predicted value of the sensor information at time t; the single-step prediction model takes the prediction sequence of each sensor at time t as input and takes the predicted value of the sensor information of each sensor at time t as output;
[0009] Determine whether the sensor is abnormal according to the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t;
[0010] If the sensor is not abnormal, the sensor information sampling value of the sensor at time t is added to the end of the last value of the prediction sequence of the sensor at time t, and the first value of the prediction sequence of the sensor at time t is removed to obtain the prediction sequence of the sensor at time t+1, and time t+1 is taken as time t, and the step of "using the prediction sequence of the sensor at time t as input and using the single-step prediction model to determine the prediction value of the sensor information at time t" is returned;
[0011] If the sensor is abnormal, perform the following steps:
[0012] The sensor information sampling value of the sensor at time t is added to the end of the last value of the prediction sequence of the sensor at time t, and the first value of the prediction sequence of the sensor at time t is removed to obtain the prediction sequence of the sensor at time t+1;
[0013] Taking the prediction sequence of the sensor at time t+1 as input, the single-step prediction model is used to determine the predicted value of the sensor information at time t+1;
[0014] The sensor information prediction value of the sensor at time t+1 is added to the end of the last value of the prediction sequence of the sensor at time t+1, and the first value of the prediction sequence of the sensor at time t+1 is removed to obtain the prediction sequence of the sensor at time t+2;
[0015] Taking the prediction sequence of the sensor at time t+2 as input, the single-step prediction model is used to determine the predicted value of the sensor information at time t+2;
[0016] Determine whether the abnormality of the sensor has ended according to the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2;
[0017] If the sensor anomaly has not ended, the predicted value of the sensor information at time t+2 is added to the end of the last value of the prediction sequence of the sensor at time t+2, and the first value of the prediction sequence of the sensor at time t+2 is removed to obtain the prediction sequence of the sensor at time t+3; time t+3 is taken as time t+2, and the step of "using the prediction sequence of the sensor at time t+2 as input and using the single-step prediction model to determine the predicted value of the sensor information at time t+2" is returned;
[0018] If the sensor ends abnormally, the sensor information prediction value of the sensor at time t+2 is added to the back of the last value of the prediction sequence of the sensor at time t+2, and the first value of the prediction sequence of the sensor at time t+2 is removed to obtain the prediction sequence of the sensor at time t+3; the prediction sequence of the sensor at time t+3 is used as input, and the sensor information prediction value of the sensor at time t+3 is determined by using the single-step prediction model; the sensor information sampling value of the sensor at time t+3 is added to the back of the last value of the prediction sequence of the sensor at time t+3, and the first value of the prediction sequence of the sensor at time t+3 is removed to obtain the prediction sequence of the sensor at time t+4; take time t+4 as time t, and return to the step of "using the prediction sequence of the sensor at time t as input, and determining the sensor information prediction value of the sensor at time t by using the single-step prediction model".
[0019] In some embodiments, the single-step prediction model includes a graph learning module and several prediction modules connected in sequence, and the prediction module includes a timing analysis module and a graph convolution module connected in sequence.
[0020] In some embodiments, determining whether an abnormality occurs in the sensor according to the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t specifically includes:
[0021] Calculate the mean square error between the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t to obtain a first monitoring indicator;
[0022] Determine whether the first monitoring indicator reaches a first abnormal threshold; if so, determine that the sensor is abnormal; if not, determine that the sensor is not abnormal.
[0023] In some embodiments, if the sensor is abnormal, the method further includes:
[0024] Determine whether the first monitoring indicator of the sensor reaches a second abnormal threshold value, and obtain a first determination result;
[0025] If the first judgment result is no, then the abnormality type of the sensor is abnormal sensing position;
[0026] If the first judgment result is yes, then the difference between the sensor information sampling value of the sensor at time t and the sensor information sampling value of the sensor at time t-1 is calculated to obtain a second monitoring indicator; and whether the second monitoring indicator has been 0 for multiple consecutive moments to obtain a second judgment result;
[0027] If the second judgment result is yes, the abnormality type of the sensor is communication abnormality;
[0028] If the second judgment result is no, then the abnormality type of the sensor is environmental interference abnormality.
[0029] In some embodiments, after determining that the abnormality type of the sensor is an environmental interference abnormality, the method further includes:
[0030] Calculate the difference between the sensor information sampling value of the sensor at time t and the sensor information predicted value of the sensor at time t to obtain a third monitoring index; determine whether the third monitoring index is greater than 0; if so, the abnormal type of the sensor is cutting heat intervention; if not, the abnormal type of the sensor is coolant intervention.
[0031] In some embodiments, after determining the abnormal type of the sensor, the method further includes:
[0032] Determine whether the number of sensors with the same abnormal type exceeds N / 2; if so, the complex equipment may have a fault; where N is the total number of sensors.
[0033] In some embodiments, the abnormal sensing position includes abnormal falling off, abnormal displacement and poor contact.
[0034] In some embodiments, determining whether the abnormality of the sensor has ended according to the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2 specifically includes:
[0035] Calculate the mean square error between the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2 to obtain the abnormal information correction index;
[0036] Determine whether the abnormal information correction index reaches a third abnormal threshold; if so, determine that the abnormality of the sensor has not ended; if not, determine that the abnormality of the sensor has ended.
[0037] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0038] The present invention provides an online identification and self-calibration method for abnormal sensor information in the service process of equipment. The single-step prediction model is used to determine the sensor information prediction value of each sensor. Whether the sensor is abnormal is determined based on the sensor information sampling value and the sensor information prediction value, thereby realizing online identification of abnormal sensor information. Meanwhile, if the sensor is abnormal, the sensor information prediction value of the sensor is used to replace the sensor information sampling value to monitor the service performance of complex equipment. Meanwhile, the sensor information prediction value of the sensor at the next moment is predicted, thereby realizing self-calibration of abnormal sensor information and ensuring the stability of monitoring the service process of complex equipment and the quality of sensor information. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0040] Figure 1 A method flow chart of an online identification and self-calibration method of abnormal sensor information for equipment service process provided in Example 1 of the present invention.
[0041] Figure 2 A schematic diagram of the structure of a multi-variable sensor graph network provided in Example 1 of the present invention.
[0042] Figure 3 A schematic diagram of abnormal sensor information monitoring provided in Example 1 of the present invention; Figure 3 (a) is a monitoring diagram of abnormal shedding and abnormal displacement; Figure 3 (b) is a monitoring diagram of coolant intervention; Figure 3 (c) is a monitoring diagram of poor contact and abnormal communication; Figure 3 (d) is a schematic diagram for monitoring cutting heat intervention.
[0043] Figure 4 A schematic diagram of the online identification process of various abnormal sensing patterns provided in Example 1 of the present invention.
[0044] Figure 5 This is a schematic diagram of the self-calibration principle of abnormal sensing information provided in Example 1 of the present invention. DETAILED DESCRIPTION
[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0046] The purpose of the present invention is to provide an online identification and self-calibration method for abnormal sensor information in the service process of equipment, which can realize online accurate identification and self-calibration of abnormal sensor information, and ensure the stability of monitoring the service process of complex equipment and the quality of sensor information.
[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0048] Example 1
[0049] like Figure 1 As shown, in this embodiment, an abnormal sensor information online identification and self-calibration method for the equipment service process is performed for each sensor arranged on the complex equipment. The following steps are performed:
[0050] (1) The sensor information sampling values of the sensor at multiple consecutive historical moments before time t are selected to form a prediction sequence of the sensor at time t; the sensor is used to monitor the service performance of complex equipment; the last moment of the multiple consecutive historical moments is time t-1.
[0051] (2) Taking the prediction sequence of the sensor at time t as input, a single-step prediction model is used to determine the predicted value of the sensor information at time t; the single-step prediction model takes the prediction sequence of each sensor at time t as input, and takes the predicted value of the sensor information of each sensor at time t as output.
[0052] (3) Determine whether an abnormality occurs in the sensor based on the sensor information sampling value at time t and the sensor information prediction value at time t.
[0053] (4) If no abnormality occurs in the sensor, the sensor information sampling value of the sensor at time t is added to the end of the last value of the prediction sequence of the sensor at time t, and the first value of the prediction sequence of the sensor at time t is removed to obtain the prediction sequence of the sensor at time t+1. Time t+1 is taken as time t, and the step of "using the prediction sequence of the sensor at time t as input and using the single-step prediction model to determine the prediction value of the sensor information at time t" is returned.
[0054] (5) If the sensor fails, perform the following steps:
[0055] (5.1) The sensor information sampling value of the sensor at time t is added to the end of the last value of the prediction sequence of the sensor at time t, and the first value of the prediction sequence of the sensor at time t is removed to obtain the prediction sequence of the sensor at time t+1.
[0056] (5.2) Taking the prediction sequence of the sensor at time t+1 as input, the single-step prediction model is used to determine the predicted value of the sensor information at time t+1.
[0057] (5.3) The predicted value of the sensor information at time t+1 is added to the end of the last value of the prediction sequence of the sensor at time t+1, and the first value of the prediction sequence of the sensor at time t+1 is removed to obtain the prediction sequence of the sensor at time t+2.
[0058] (5.4) Taking the prediction sequence of the sensor at time t+2 as input, the single-step prediction model is used to determine the predicted value of the sensor information at time t+2.
[0059] (5.5) Determine whether the abnormality of the sensor has ended based on the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2.
[0060] (5.6) If the sensor anomaly has not ended, the predicted value of the sensor information at time t+2 is added to the end of the last value of the prediction sequence of the sensor at time t+2, and the first value of the prediction sequence of the sensor at time t+2 is removed to obtain the prediction sequence of the sensor at time t+3; time t+3 is taken as time t+2, and the step of "using the prediction sequence of the sensor at time t+2 as input and using the single-step prediction model to determine the predicted value of the sensor information at time t+2" is returned.
[0061] (5.7) If the sensor ends abnormally, the predicted value of the sensor information at time t+2 is added to the back of the last value of the prediction sequence of the sensor at time t+2, and the first value of the prediction sequence of the sensor at time t+2 is removed to obtain the prediction sequence of the sensor at time t+3; the prediction sequence of the sensor at time t+3 is used as input, and the predicted value of the sensor information at time t+3 is determined by using the single-step prediction model; the sensor information sampling value of the sensor at time t+3 is added to the back of the last value of the prediction sequence of the sensor at time t+3, and the first value of the prediction sequence of the sensor at time t+3 is removed to obtain the prediction sequence of the sensor at time t+4; time t+4 is used as time t, and the step of "using the prediction sequence of the sensor at time t as input and determining the predicted value of the sensor information at time t by using the single-step prediction model" is returned.
[0062] In this embodiment, the complex equipment may be a CNC machine tool, a nuclear power equipment, a gas turbine, etc. All sensors may be evenly arranged on the complex equipment to collect the operating parameters of the complex equipment. The sensor information sampling value of the sensor at the current moment is the operating parameter of the complex equipment collected by the sensor at the current moment, so as to monitor the service performance of the complex equipment. The type of sensor depends on the user's needs and can be the same or different. When the complex equipment is a CNC machine tool, the sensor may be a temperature sensor; when the complex equipment is a nuclear power equipment, the sensor may be a flow rate sensor; when the complex equipment is a gas turbine, the sensor may be a pressure sensor.
[0063] The sensors of this embodiment can be evenly arranged on the core components of complex equipment to monitor the service performance of the core components of complex equipment. The core components are components that affect the core working objectives of complex equipment. For example, for CNC machine tools, the core components can be the spindle, for nuclear power equipment, the core components can be the reactor pressure vessel, and for gas turbines, the core components can be the compressor and turbine blades. It should be noted that the core components can be selected according to user needs.
[0064] like Figure 2 As shown, the single-step prediction model takes the prediction sequence of each sensor at time t as input and the predicted value of the sensor information of each sensor at time t as output. The input dimension of the single-step prediction model is seq-in-len*N, and the output dimension of the single-step prediction model is 1*N, where seq-in-len is the number of moments included in the prediction sequence and N is the number of sensors. The single-step prediction model of this embodiment adopts a multi-variable sensor graph network, which is composed of a graph learning module, a plurality of timing analysis modules, and a plurality of graph convolution modules, that is, the prediction model includes a graph learning module and a plurality of prediction modules connected in sequence, and the prediction module includes a timing analysis module and a graph convolution module connected in sequence. Figure 2 The diagram shows a prediction module. The graph learning module adaptively captures the spatial relationship between different sensors by establishing an adjacency matrix. The time series analysis module uses the Transformer model structure and considers the correlation in the prediction sequence of each sensor to better capture the long-range dependency of sensor information. The graph convolution module contains information propagation and information selection mechanisms to integrate the sensor information of each sensor and its associated sensors.
[0065] More specifically, the graph learning module adaptively learns a graph adjacency matrix to capture the hidden relationships between time series data. The adjacency matrix of the graph structure is calculated using the following formula:
[0066] M 1 =tanh(αE 1 Θ 1 ) (1)
[0067] M 2 =tanh(αE 2 Θ 2 ) (2)
[0068]
[0069] idx=argtopk(A[i,:]) (4)
[0070] A[i,-idx]=0 (5)
[0071] In formula (1), formula (2), formula (3), formula (4), and formula (5), α is a hyperparameter used to control the saturation rate of the activation function; E 1 、E 2 represents randomly initialized node embedding, each node represents a sensor, E 1 、E 2 are all matrices with N rows and P columns, where N is the total number of sensors, P is an arbitrary constant, and E 1 、E 2 Can be learned during training; 1 , Θ 2 are model parameters; A is the adjacency matrix; idx is the associated sensor of the ith sensor, i = 1, 2, ..., N; argtopk() means returning the top-k index, that is, returning the sensors corresponding to the top k elements in the ith row of the adjacency matrix A when the elements are arranged in descending order; -idx represents other sensors except the associated sensors of the ith sensor.
[0072] The asymmetric property of the graph adjacency matrix proposed in this embodiment is realized by formula (3), and the subtraction term and ReLU activation function are used to regularize the adjacency matrix. Formula (4) and formula (5) are a strategy to make the adjacency matrix sparse, while reducing the computational cost of the subsequent graph convolution module. For each node (i.e., sensor), its top-k nearest nodes are selected as its neighbor nodes (i.e., associated sensors), and the weights of non-neighbor nodes are set to zero while retaining the weights of neighbor nodes.
[0073] The time series analysis module uses the Transformer model. The input is a time series (i.e., the sequence used for prediction), and the output is a time series that captures long-range dependencies.
[0074] The graph convolution module uses a hybrid skip layer. The input is the adjacency matrix and a time series that captures long-range dependencies, and the output is the output result after graph convolution.
[0075] If multiple prediction modules are included, that is, multiple timing analysis modules and multiple graph convolution modules, the output result of the previous prediction module after graph convolution is used as the input of the current prediction module (that is, the prediction sequence).
[0076] For each prediction module, this embodiment introduces jump connections after the timing analysis module and the graph convolution module respectively, and inputs the time series that captures the long-range dependency and the output result after the graph convolution into the jump connection layer, and then sends them to the output layer through the jump connection layer. The output layer integrates the information of all jump connection layers to obtain the prediction value of the sensor information.
[0077] In this embodiment, a physical prototype test bench for core components of complex equipment is pre-built, and N key sensors are evenly arranged on the physical prototype test bench for core components of complex equipment to monitor the service performance of core components of complex equipment. The sensor information sampling values obtained by N key sensors during the operation of the physical prototype test bench for core components of complex equipment are collected, and the sensor information sampling values at multiple consecutive moments are used as input, and the sensor information sampling values at the next moment of multiple consecutive moments are used as output. The multi-variable sensor graph network is trained to establish a high-precision single-step prediction model (i.e., a single-step prediction model) for N-dimensional sensor information. Subsequently, the prediction sequence of each sensor at the current moment is used as input, and the prediction model is used to determine the sensor information prediction value of each sensor at the current moment.
[0078] Considering that current online monitoring methods generally have the problem of weak abnormal sensor pattern recognition ability and lack of accurate sensor abnormality source identification ability, this embodiment conducts physical simulation tests that may cause various sensor information abnormal patterns (i.e., abnormal sensor patterns, also known as abnormal types) on a physical prototype test bench of a core component of complex equipment based on the actual service conditions of complex equipment. The sensor information abnormal patterns that may exist under the actual service conditions of complex equipment include: abnormal sensor position (abnormal detachment, abnormal displacement, poor contact), communication abnormality, environmental interference (temperature change), etc. Environmental interference includes cutting heat interference and coolant interference. Cutting heat interference refers to the interference caused to the sensor by the contact between the cutting heat applied to the complex equipment and the sensor. Coolant interference refers to the interference caused to the sensor by the contact between the coolant applied to the complex equipment and the sensor. When conducting the physical simulation test of each sensor information abnormal pattern, the sensor information sampling value of the i-th sensor (i=1,2,…N) at each moment is calculated. The predicted value of the sensor information of the i-th sensor (i=1,2,…N) at the corresponding time obtained by the high-precision single-step prediction model Compare and obtain the abnormal information monitoring indicators of the abnormal mode of the sensor information, such as Figure 3 As shown. The sensor information representation forms corresponding to various sensor information abnormal modes are analyzed. When the current sensor information sampling value is judged to be abnormal, the current sensor information abnormal mode is identified according to the abnormal information monitoring index, thereby establishing a relationship between the abnormal information monitoring index and the sensor information abnormal mode. Subsequently, the sensor information abnormal mode can be determined based on the abnormal information monitoring index, thereby having the ability to accurately identify the sensor abnormal source.
[0079] The abnormal information monitoring index includes a first monitoring index, a second monitoring index and a third monitoring index. In this embodiment, the sensor information sampling value of the i-th sensor at time t and the predicted value of the sensor information of the i-th sensor at time t The mean square error between the two is used as the first monitoring indicator, with the sensor information sampling values of the i-th sensor at adjacent times t and t-1 The difference between the two is used as the second monitoring indicator, and the sensor information sampling value of the i-th sensor at time t is and the predicted value of the sensor information of the i-th sensor at time t The difference between them is taken as the third monitoring indicator.
[0080] The first monitoring indicator is as follows:
[0081]
[0082] In formula (6), Index_1 is the first monitoring index; is the sensor information sampling value of the i-th sensor at time t; is the predicted value of the sensor information of the i-th sensor at time t.
[0083] The second monitoring indicator is as follows:
[0084]
[0085] In formula (7), Index_2 is the second monitoring index; is the sensor information sampling value of the i-th sensor at time t-1.
[0086] The third monitoring indicator is as follows:
[0087]
[0088] In formula (8), Index_3 is the third monitoring index.
[0089] like Figure 4 As shown, the first monitoring indicator Index_1 reaches the first abnormal threshold as the judgment standard for abnormal sensor information. If it reaches the first abnormal threshold, the sensor is abnormal. If it does not reach the first abnormal threshold, the sensor is not abnormal. If the Index_1 values of not less than (i.e., more than) N / 2 sensors are simultaneously judged to be abnormal at the same level (i.e., the same abnormal type), it means that the sensor itself may not be abnormal, and it is believed that the core components of the complex equipment may be faulty. If the Index_1 values of less than (i.e., less than) N / 2 sensors are simultaneously judged to be abnormal at the same level, it means that the sensor itself is abnormal. For each abnormal sensor, the following judgment is made: If the Index_1 value of the abnormal sensor does not exceed (i.e., does not reach) the second abnormal threshold, it is judged as a sensor position abnormality (abnormal detachment, abnormal displacement, poor contact); If the Index_1 value of the abnormal sensor exceeds (i.e., reaches) the second abnormal threshold, the second monitoring indicator Index_2 is activated. If several consecutive Index_2 values of the abnormal sensor are all 0, it is judged as a communication abnormality; If several consecutive Index_2 values of the abnormal sensor are not all 0, the third monitoring indicator Index_3 is activated. Based on the Index_2 value of the abnormal sensor, the third monitoring indicator Index_4 is activated. 3 The value is greater than or less than 0 to determine the specific form of environmental interference. If an abnormal sensor index 3 If the value is greater than 0, it is judged as cutting heat interference. If the index of the sensor where the abnormality occurs is 3 If the value is not greater than 0, it is judged as coolant intervention.
[0090] In this embodiment, determining whether the sensor is abnormal according to the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t specifically includes:
[0091] (1) Calculate the mean square error between the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t to obtain a first monitoring index.
[0092] (2) Determine whether the first monitoring indicator reaches a first abnormal threshold; if so, determine that the sensor is abnormal; if not, determine that the sensor is not abnormal.
[0093] If an abnormality occurs in the sensor, the method of this embodiment also includes: determining whether the first monitoring indicator of the sensor reaches a second abnormal threshold value to obtain a first judgment result; if the first judgment result is no, the abnormality type of the sensor is a sensor position abnormality; if the first judgment result is yes, calculating the difference between the sensor information sampling value of the sensor at time t and the sensor information sampling value of the sensor at time t-1 to obtain a second monitoring indicator; determining whether the second monitoring indicator has been 0 for multiple consecutive moments to obtain a second judgment result; if the second judgment result is yes, the abnormality type of the sensor is a communication abnormality; if the second judgment result is no, the abnormality type of the sensor is an environmental interference abnormality.
[0094] After determining that the abnormality type of the sensor is an environmental interference abnormality, the method of this embodiment also includes: calculating the difference between the sensor information sampling value of the sensor at time t and the sensor information predicted value of the sensor at time t to obtain a third monitoring indicator; judging whether the third monitoring indicator is greater than 0; if so, the abnormality type of the sensor is cutting heat interference; if not, the abnormality type of the sensor is coolant interference.
[0095] After determining the abnormal type of the sensor, the method of this embodiment further includes: determining whether the number of sensors with the same abnormal type exceeds N / 2; if so, the complex equipment may have a fault; wherein N is the total number of sensors.
[0096] like Figure 5 As shown, the self-calibration principle of the abnormal sensor information online identification and self-calibration method of this embodiment is: if the i-th sensor is abnormal at time t, the sensor information prediction value of the i-th sensor at time t+1 obtained by the high-precision single-step prediction model when the i-th sensor is abnormal Replace the sensor information sampling value of the i-th sensor at time t+1 To monitor the service performance of core components of complex equipment, and As the calibration input of the high-precision single-step prediction model, it continuously outputs the calibrated sensor information prediction value Until the anomaly disappears at time t+m, when the anomaly disappears, there is no need to replace it, and the sensor information sampling value of the i-th sensor at time t+m+1 is directly reused To monitor the service performance of core components of complex equipment, Re-input into the high-precision single-step prediction model to identify and calibrate abnormal sensor information online.
[0097] This embodiment uses the abnormal information correction index to determine whether the abnormal situation has disappeared. The abnormal information correction index is as follows:
[0098]
[0099] In formula (9), Index_4 is the abnormal information correction index; is the sensor information sampling value of the i-th sensor at time t+m; is the predicted value of the sensor information of the i-th sensor at time t+m.
[0100] The abnormal information correction index Index_4 not exceeding the third abnormal threshold is used as a criterion for determining whether the abnormality of the i-th sensor disappears.
[0101] In this embodiment, based on the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2, it is determined whether the sensor abnormality has ended, specifically including: calculating the mean square error between the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2 to obtain the abnormal information correction index; judging whether the abnormal information correction index reaches the third abnormal threshold; if so, determining that the sensor abnormality has not ended; if not, determining that the sensor abnormality has ended.
[0102] The present embodiment provides an online identification and self-calibration method for abnormal sensor information during the service process of equipment. The method builds a physical prototype test bench for core components of complex equipment, collects sensor information during the operation of the physical prototype test bench for core components of complex equipment, and uses a variable sensor graph network to establish a high-precision single-step prediction model for multi-dimensional sensor information, simulates and analyzes various abnormal sensor modes and sensor information representation forms of the service conditions of complex equipment, uses abnormal information monitoring indicators to monitor and identify the current abnormal sensor mode, replaces the sensor information sampling value with the sensor information prediction value during the abnormal period, and inputs the sensor information sampling value into the high-precision single-step prediction model to calibrate the abnormal information online, and continuously cycles until the abnormality disappears, and reuses the sensor information sampling value to monitor the service performance of the core component, continues to identify and calibrate the sensor abnormality online, and monitors the sensor abnormality online through the online identification and self-calibration method, thereby ensuring the stability of the monitoring of the service process of the complex equipment and the quality of the sensor information.
[0103] Compared with the prior art, this embodiment has the following beneficial effects:
[0104] (1) By establishing a high-precision single-step prediction model and complete abnormal information monitoring indicators, accurate online identification of abnormal sensor information of complex equipment can be achieved, which helps to timely discover and deal with abnormal situations.
[0105] (2) Under abnormal circumstances, the sensor information prediction value is used instead of the sensor information sampling value to input into the high-precision single-step prediction model for sensor information cyclic self-calibration, ensuring the stability of complex equipment service process monitoring and the quality of sensor information.
[0106] Hereinafter, this embodiment takes a complex equipment as a CNC machine tool as an example to further introduce the online recognition and self-calibration method of abnormal sensor information of this embodiment:
[0107] (1) Build a physical prototype test bench for core components of complex equipment and evenly arrange N key sensors to monitor the service performance of core components of complex equipment.
[0108] Specifically, a physical prototype test bench for the spindle of a precision CNC machine tool was built, and 11 temperature sensors were evenly arranged near the spindle to monitor the thermal characteristics of the spindle during operation.
[0109] (2) The sensor information of the test bench during operation is collected, and a high-precision single-step prediction model of N-dimensional sensor information is established using a multi-variable sensor graph network.
[0110] Specifically, the temperature sensor information during the operation of the spindle physical prototype test bench is collected, and a high-precision single-step prediction model of 11-dimensional temperature sensor information is established using a multi-variable sensor graph network. Among them, the multi-variable sensor graph network is mainly composed of a graph learning module, several timing analysis modules, and several graph convolution modules. The graph learning module adaptively captures the spatial relationship between different temperature sensors by establishing an adjacency matrix. This adaptability enables the multi-variable sensor graph network to flexibly adapt to different environments, better understand the positional relationship between temperature sensors, and help to more accurately predict temperature changes. The timing analysis module uses the Transformer model structure to simultaneously consider the correlation in the sampling sequence of each temperature sensor, thereby better capturing the long-range dependency of temperature sensor information. Relying on the powerful timing analysis capability of the Transformer model, the multi-variable sensor graph network can model the evolution trend of temperature sensor information in the time dimension, thereby better predicting future temperature changes. The graph convolution module contains information propagation and information selection mechanisms, which are used to integrate the sensor information of each temperature sensor and its associated temperature sensors. Through the graph convolution operation, the multi-variable sensor graph network can effectively aggregate the information of neighboring temperature sensors and select according to the importance of temperature information to improve the accuracy and robustness of temperature prediction. The output dimension value of the output layer is 1, and a high-precision single-step prediction task is performed.
[0111] In addition, the introduction of the jump connection layer is used to normalize the information, unify the dimension of the information spliced from multiple modules and send it to the output layer, which enhances the fluidity of information and avoids information loss.
[0112] (3) According to the actual service conditions of complex equipment, physical simulation tests that may cause various abnormalities in sensor information are carried out on the test bench, and the sensor information sampling values of the i-th sensor (i = 1, 2, ... N) at each moment are calculated. The predicted value of the sensor information at the corresponding moment predicted by the high-precision single-step prediction model Perform comparisons to obtain abnormal information monitoring indicators.
[0113] Specifically, the actual service conditions of CNC machine tools are relatively harsh, usually accompanied by cutting heat and coolant splashing. Once the temperature sensor accidentally comes into contact with cutting heat and coolant, it will inevitably cause abnormal temperature changes of the temperature sensor. In addition, due to accidents or human intervention, there is also the possibility of abnormal position of the temperature sensor, which is specifically manifested in the form of: sensor detachment, sensor position change, poor contact between the sensor and complex equipment, etc. Sometimes, communication failures between sensors, data acquisition cards and computers are also one of the reasons for abnormal sensor information.
[0114] Various abnormal sensing modes are simulated on the spindle physical prototype test bench, and the sensing information sampling value of the i-th sensor (i = 1, 2, ... N) at each moment is used. Prediction value of sensor information The mean square error between the two is used as the first monitoring indicator; the sensor information sampling value of the i-th sensor at adjacent moments The difference between them is used as the second monitoring indicator; the sensor information sampling value of the i-th sensor at each moment Prediction value of sensor information The difference between them is taken as the third monitoring indicator.
[0115] The first monitoring indicator is as follows:
[0116]
[0117] The second monitoring indicator is as follows:
[0118]
[0119] The third monitoring indicator is as follows:
[0120]
[0121] (4) Analyze the sensor information representation forms corresponding to various abnormal sensor modes, and when the current abnormal sensor information is judged to be abnormal, identify the current abnormal sensor mode according to the abnormal information monitoring indicators.
[0122] Specifically, for various abnormal sensing modes (sensing position abnormality, coolant intervention, cutting heat intervention, communication abnormality, etc.), the first monitoring indicator Index_1 reaching the first abnormal threshold is used as the judgment standard for abnormal sensing information (such as Figure 3 As shown), if the Index_1 values of no less than N / 2 sensors are judged to be abnormal at the same level at the same time, it is considered that the core components of the complex equipment may have faults. When the Index_1 values of no more than N / 2 sensors are judged to be abnormal at the same level at the same time, the following judgment is made: If the Index_1 value does not exceed the second abnormal threshold, it is judged to be abnormal in the sensor position (abnormal detachment, abnormal displacement, poor contact); if the Index_1 value exceeds the second abnormal threshold, the second monitoring indicator Index_2 is activated. If several consecutive Index_2 values are 0, it is judged to be a communication abnormality; if the Index_2 value is not 0, the third monitoring indicator Index_3 is activated, and Index_4 is used as the monitoring indicator. 3 The value is greater than or less than 0 to determine the specific form of environmental interference abnormality. The abnormal sensing mode online identification flow chart is as follows: Figure 4 shown.
[0123] (5) The predicted value of the sensor information at time t+1 obtained by the high-precision single-step prediction model when the i-th sensor is abnormal (at time t) Replace the sensor information sampling value at time t+1 Monitor the service performance of core components and use it as the calibration input for the high-precision single-step prediction model, and continue to cycle until the anomaly disappears at time t+m, and reuse the sensor information sampling value While monitoring the service performance of core components, abnormal sensor information can be identified and calibrated online.
[0124] Specifically, when the i-th temperature sensor is abnormal at time t, the predicted value of the temperature sensor information at time t+1 obtained by the high-precision single-step prediction model is Replace the temperature sensing information sampling value at time t+1 As the calibration input of the high-precision single-step prediction model, the calibrated temperature sensor information prediction value is continuously output in a loop The abnormal information correction index is used to determine whether the abnormal situation has disappeared. After the abnormality disappears at time t+m, the temperature sensor information sampling value is used again to monitor the service performance of the core components. Re-input into the high-precision single-step prediction model to identify and calibrate abnormal temperature sensor information online.
[0125] The abnormal information correction indicators are as follows:
[0126]
[0127] The abnormal information correction index Index_4 not exceeding the third abnormal threshold is used as a criterion for determining whether the abnormality of the i-th temperature sensor disappears.
[0128] This embodiment proposes an online identification and self-calibration method for abnormal sensor information in the service process of equipment, establishes a high-precision single-step prediction model for key sensor information of complex equipment, simulates various sensor anomalies and uses abnormal information monitoring indicators to identify the current abnormal mode, and replaces the sampled values with predicted values and inputs them into the high-precision single-step prediction model to calibrate the abnormal sensor information, thereby ensuring the stability of monitoring the service process of complex equipment and the quality of sensor information.
[0129] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. An online identification and self-calibration method for abnormal sensor information during equipment service. It is characterized in that For each sensor deployed on the complex, perform the following steps: The sensor information sampling values of the sensor at multiple consecutive historical moments before time t are selected to form a prediction sequence of the sensor at time t; the sensor is used to monitor the service performance of complex equipment; the last moment of the multiple consecutive historical moments is time t-1; Taking the prediction sequence of the sensor at time t as input, using a single-step prediction model to determine the predicted value of the sensor information at time t; the single-step prediction model takes the prediction sequence of each sensor at time t as input and takes the predicted value of the sensor information of each sensor at time t as output; Determine whether the sensor is abnormal according to the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t; If the sensor is not abnormal, the sensor information sampling value of the sensor at time t is added to the end of the last value of the prediction sequence of the sensor at time t, and the first value of the prediction sequence of the sensor at time t is removed to obtain the prediction sequence of the sensor at time t+1, and time t+1 is taken as time t, and the step of "using the prediction sequence of the sensor at time t as input and using the single-step prediction model to determine the prediction value of the sensor information at time t" is returned; If the sensor is abnormal, perform the following steps: The sensor information sampling value of the sensor at time t is added to the end of the last value of the prediction sequence of the sensor at time t, and the first value of the prediction sequence of the sensor at time t is removed to obtain the prediction sequence of the sensor at time t+1; Taking the prediction sequence of the sensor at time t+1 as input, the single-step prediction model is used to determine the predicted value of the sensor information at time t+1; The sensor information prediction value of the sensor at time t+1 is added to the end of the last value of the prediction sequence of the sensor at time t+1, and the first value of the prediction sequence of the sensor at time t+1 is removed to obtain the prediction sequence of the sensor at time t+2; Taking the prediction sequence of the sensor at time t+2 as input, the single-step prediction model is used to determine the predicted value of the sensor information at time t+2; Determine whether the abnormality of the sensor has ended according to the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2; If the sensor anomaly has not ended, the predicted value of the sensor information at time t+2 is added to the end of the last value of the prediction sequence of the sensor at time t+2, and the first value of the prediction sequence of the sensor at time t+2 is removed to obtain the prediction sequence of the sensor at time t+3; time t+3 is taken as time t+2, and the step of "using the prediction sequence of the sensor at time t+2 as input and using the single-step prediction model to determine the predicted value of the sensor information at time t+2" is returned; If the sensor ends abnormally, the predicted value of the sensor information at time t+2 is added to the back of the last value of the prediction sequence of the sensor at time t+2, and the first value of the prediction sequence of the sensor at time t+2 is removed to obtain the prediction sequence of the sensor at time t+3; the prediction sequence of the sensor at time t+3 is used as input, and the predicted value of the sensor information at time t+3 is determined by using the single-step prediction model; the sensor information sampling value of the sensor at time t+3 is added to the back of the last value of the prediction sequence of the sensor at time t+3, and the first value of the prediction sequence of the sensor at time t+3 is removed to obtain the prediction sequence of the sensor at time t+4; time t+4 is used as time t, and the step of "using the prediction sequence of the sensor at time t as input, and determining the predicted value of the sensor information at time t by using the single-step prediction model" is returned.
2. According to claim 1, a method for online identification and self-calibration of abnormal sensor information for equipment service process, It is characterized in that The single-step prediction model includes a graph learning module and several prediction modules connected in sequence, and the prediction module includes a timing analysis module and a graph convolution module connected in sequence.
3. According to claim 1, a method for online identification and self-calibration of abnormal sensor information for equipment service process, It is characterized in that According to the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t, it is determined whether the sensor is abnormal, specifically including: Calculate the mean square error between the sensor information sampling value of the sensor at time t and the sensor information prediction value of the sensor at time t to obtain a first monitoring indicator; Determine whether the first monitoring indicator reaches a first abnormal threshold; if so, determine that the sensor is abnormal; if not, determine that the sensor is not abnormal.
4. According to claim 3, a method for online identification and self-calibration of abnormal sensor information for equipment service process, It is characterized in that If the sensor is abnormal, it also includes: Determine whether the first monitoring indicator of the sensor reaches a second abnormal threshold value, and obtain a first determination result; If the first judgment result is no, then the abnormality type of the sensor is abnormal sensing position; If the first judgment result is yes, then the difference between the sensor information sampling value of the sensor at time t and the sensor information sampling value of the sensor at time t-1 is calculated to obtain a second monitoring indicator; and whether the second monitoring indicator has been 0 for multiple consecutive moments to obtain a second judgment result; If the second judgment result is yes, the abnormality type of the sensor is communication abnormality; If the second judgment result is no, then the abnormality type of the sensor is environmental interference abnormality.
5. According to claim 4, a method for online identification and self-calibration of abnormal sensor information for equipment service process, It is characterized in that After determining that the abnormal type of the sensor is an environmental interference abnormality, the following further includes: Calculate the difference between the sensor information sampling value of the sensor at time t and the sensor information predicted value of the sensor at time t to obtain a third monitoring index; determine whether the third monitoring index is greater than 0; if so, the abnormal type of the sensor is cutting heat intervention; if not, the abnormal type of the sensor is coolant intervention.
6. According to claim 5, a method for online identification and self-calibration of abnormal sensor information for equipment service process, It is characterized in that After determining the abnormal type of the sensor, it also includes: Determine whether the number of sensors with the same abnormal type exceeds N / 2; if so, the complex equipment may have a fault; where N is the total number of sensors.
7. According to claim 4, a method for online identification and self-calibration of abnormal sensor information for equipment service process, It is characterized in that The sensor position abnormality includes abnormal falling off, abnormal displacement and poor contact.
8. According to claim 1, a method for online identification and self-calibration of abnormal sensor information for equipment service process, It is characterized in that According to the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2, it is determined whether the abnormality of the sensor has ended, specifically including: Calculate the mean square error between the sensor information sampling value of the sensor at time t+2 and the sensor information prediction value of the sensor at time t+2 to obtain the abnormal information correction index; Determine whether the abnormal information correction index reaches a third abnormal threshold; if so, determine that the abnormality of the sensor has not ended; if not, determine that the abnormality of the sensor has ended.
Citation Information
Cited By
Monitoring method and system based on Internet of Things architecture
CN119394375A
Sensor data drift online calibration method
CN119533549A
