Sensor correction method and device, storage medium and program product

Abstract

The embodiment of the invention provides a sensor correction method, a computer device, a computer readable storage medium and a computer program product. Wherein the initial automatic encoder model is trained through the collected operation parameters of the sensor in the normal state, the trained target automatic encoder model is obtained, and the target automatic encoder model is used for detecting the working state of the sensor. And then, through a mode of coupling an automatic encoder and Bayesian reasoning, carrying out equivalent sampling and iteratively updating a posterior distribution probability density function of sensor fault deviation, calculating according to equivalent sampling data to obtain a sensor target fault deviation value, and realizing online correction according to the sensor target fault deviation value. Through sensor fault on-line correction, the system control precision can be effectively improved, the operation and maintenance cost is reduced, and the service life of the sensor is prolonged.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of sensor fault diagnosis and calibration, and particularly to a sensor calibration method, a computer device, a computer-readable storage medium, and a computer program product. Background Art

[0002] The cold plate liquid cooling technology is more and more widely used in high power density scenarios such as cloud computing, big data, and 5G networks. As an important component of the cold plate liquid cooling system, the External Distribution Unit (EDU) transports the low-temperature liquid of the outdoor cold source to the data machine room through a circulating water pump to meet the heat dissipation requirements of the high heat density servers in the data center machine room. In a typical EDU control logic, the operation of the circulating water pump and the electric valve is controlled by the data collected by sensors such as pressure, temperature, and flow rate. However, due to the influence of factors such as the installation environment, daily operation, and operation and maintenance, the sensors on the EDU unit will inevitably have various faults, resulting in a decrease in the control accuracy of the cooling system. In severe cases, it may even lead to the paralysis of the cooling system and the thermal shutdown of the server.

[0003] In the related art, although the methods of sensor redundancy design and sensor reading over-range alarm can cope with this simple sensor complete failure fault, they cannot detect and identify characteristic faults such as fixed deviation, drift deviation, and accuracy degradation. Even if the fault is successfully identified, technicians are required to go to the site to replace the sensor, which is complex in operation and maintenance and consumes manpower. Therefore, how to identify various characteristic faults of sensors and realize on-line calibration and repair of sensors is a technical problem that needs to be solved urgently at present. Summary of the Invention

[0004] The embodiments of the present application provide a sensor calibration method, a computer device, a computer-readable storage medium, and a computer program product, aiming to realize the state detection of various characteristic faults of sensors and the on-line repair function of faults, improve the system control accuracy, extend the service life of sensors, and reduce the operation and maintenance costs.

[0005] First aspect, an embodiment of the present application provides a calibration method for a sensor, including: obtaining a target autoencoder model according to first operating parameters and an initial autoencoder model, where the first operating parameters at least include operating parameters in the normal state of the sensor; using the target autoencoder model to detect the operating state of the sensor to obtain a second operating state of the sensor; obtaining a first objective function of the sensor fault deviation according to second operating parameters and the initial autoencoder model, where the second operating parameters include operating parameters when the sensor is in the second operating state; establishing a first prior distribution probability density function and a multivariate normal distribution probability density function of the sensor; obtaining a first posterior distribution probability density function and a first sampling parameter sample according to the first objective function, the first prior distribution probability density function, and the multivariate normal distribution probability density function; updating the first posterior distribution probability density function according to the first sampling parameter sample to obtain a second posterior distribution probability density function; obtaining a second sampling parameter sample according to the second posterior distribution probability density function and the multivariate normal distribution probability density function; and calibrating the second operating parameters according to the first sampling parameter sample and the second sampling parameter sample.

[0006] Second aspect, an embodiment of the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor implements the calibration method for the sensor as described in the first aspect when executing the computer program.

[0007] Third aspect, an embodiment of the present application provides a computer-readable storage medium storing computer-executable instructions for executing the calibration method for the sensor as described in the first aspect.

[0008] Fourth aspect, an embodiment of the present application provides a computer program product storing computer-executable instructions for executing the calibration method for the sensor as described in the first aspect.

[0009] The embodiment of the present application diagnoses the operating state and fault location of the sensor through an autoencoder model, and realizes online calibration of the sensor fault through the coupling method of Bayesian inference and the autoencoder. The embodiment of the present application can not only identify multiple types of faults such as fixed deviation, drift deviation, accuracy degradation, and complete failure of the sensor, but also realize online repair of the fault, effectively improving the system control accuracy, promoting the automation process of the cold plate liquid cooling system. At the same time, it reduces the number of times of on-site sensor replacement by maintenance engineers, reduces the operation and maintenance cost, and extends the service life of the sensor. Description of the Drawings

[0010] Figure 1It is a schematic diagram of the application scenario of the sensor calibration method provided by an embodiment of the present application;

[0011] Figure 2 It is a flowchart of the sensor calibration method provided by an embodiment of the present application;

[0012] Figure 3 It is a schematic diagram of the autoencoder model provided by an embodiment of the present application;

[0013] Figure 4 It is a flowchart of the target autoencoder for detecting the working state of the sensor provided by an embodiment of the present application;

[0014] Figure 5 It is a flowchart of the sensor calibration method provided by another embodiment of the present application;

[0015] Figure 6 It is a flowchart of the sensor operation state detection provided by an example of the present application;

[0016] Figure 7 It is a flowchart of the sensor operation state calibration provided by an example of the present application;

[0017] Figure 8 It is a flowchart of the MCMC equivalent sampling and iteration provided by an example of the present application;

[0018] Figure 9 It is a flowchart of the sensor operation state detection and calibration provided by an example of the present application;

[0019] Figure 10 It is a schematic diagram of the structure of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0020] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

[0021] It should be noted that although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order from the module division in the device or the order in the flowchart. Terms such as "first" and "second" in the specification, claims and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence.

[0022] In the embodiments of the present application, words such as "exemplarily" and "preferably" are used to represent examples, illustrations or explanations, and should not be construed as being more preferred or having more advantages than other embodiments or design solutions. The use of words such as "exemplarily" and "preferably" aims to present relevant concepts in a specific manner.

[0023] Due to the high thermal conductivity, high heat capacity, and high latent heat of phase change of the liquid cooling working medium, cold plate liquid cooling technology is increasingly widely used in high power density scenarios such as cloud computing, big data, and 5G networks. The EDU is an important part of the cold plate liquid cooling system. It transports the low-temperature liquid from the outdoor cold source to the data computer room through a circulating water pump to meet the heat dissipation requirements of high heat density servers in the data center computer room. In a typical EDU control logic, the operation of the circulating water pump and the electric valve is controlled by the data collected by sensors such as pressure, temperature, and flow rate. However, due to the influence of factors such as the installation environment, daily operation, and operation and maintenance, various characteristic faults are inevitable in the sensors on the EDU unit, such as complete failure, fixed fault deviation, drift fault deviation, and accuracy degradation. These faults will cause a decrease in the control accuracy of the cooling system, and in severe cases, it may even lead to the paralysis of the cooling system and the thermal shutdown of the server.

[0024] In the related art, although the sensor redundancy design and the method of warning when the sensor reading exceeds the range can cope with the complete failure of the sensor, it is still impossible to detect and identify faults such as fixed deviation faults, drift deviation faults, and accuracy degradation. Even if the fault is successfully identified, it still requires maintenance personnel to go to the site to replace the sensor.

[0025] Based on this, the embodiments of the present application provide a method for online calibration of sensors, which realizes online detection and calibration of sensor faults. The working state and fault location of the sensor are diagnosed through an autoencoder, and online calibration of the sensor fault is realized through the coupling of Bayesian inference and the autoencoder, solving the problems of complex control source fault characteristics, difficult identification, and low operation efficiency of the internal pressure, temperature, and flow sensors of the power equipment in the cold plate liquid cooling system of the existing data center computer room. The embodiments of the present application do not need to establish an accurate physical mechanism model, significantly reducing the cognitive requirements of maintenance personnel for the internal association of the cooling system. Through the mutual feedback of the autoencoder and Bayesian inference, online calibration is realized, greatly improving the service life of the sensor and reducing the energy consumption of the liquid cooling system operation.

[0026] The sensor calibration method provided by the embodiments of the present application can be applied to various power density scenarios such as data computer rooms. Figure 1 It is a schematic diagram of the application scenario of the sensor calibration method provided by an embodiment of the present application. Figure 1 It shows the EDU system process of the cold plate liquid cooling system, as Figure 1As shown, the liquid outlet and inlet are referenced to the plate heat exchanger. Labels 1 to 4 represent the sensors installed at the liquid outlet, where 1 and 2 are liquid outlet pressure sensors, and 3 and 4 are liquid outlet temperature sensors. The liquid outlet is sequentially connected to a water pump 6 and a check valve 9. A water pump inlet shock-absorbing hose 5 and a water pump outlet shock-absorbing hose 7 are respectively installed at the inlet and outlet of the water pump 6. A water pump outlet pressure sensor 8 is installed between the water pump 6 and the check valve 9. After the liquid comes out of the check valve 9, it flows towards the filter 11 through the cold source side 17. A pre-filter pressure sensor 10 and a flow sensor 12 are respectively installed at the inlet and outlet of the filter 11. After the liquid is filtered by the filter 11, it flows towards the plate heat exchanger side 18. An inlet liquid temperature sensor 13, an inlet liquid temperature sensor 14, an inlet liquid pressure sensor 15 and an inlet liquid pressure sensor 16 are also installed between the filter 11 and the plate heat exchanger side 18.

[0027] To elaborate on this technical solution in detail, the calibration method of the sensors in the embodiments of this application is further described.

[0028] Figure 2 It is a flowchart of a calibration method for sensors provided by an embodiment of this application. As Figure 2 shown, the calibration method of the sensors includes but is not limited to S1000 to S1700.

[0029] S1000: Obtain a target autoencoder model according to the first operating parameter and the initial autoencoder model, where the first operating parameter includes at least the operating parameters in the normal state of the sensor.

[0030] In one embodiment, obtain the first operating parameter, and use the training sample data provided by the first operating parameter to train the initial autoencoder model to obtain the target autoencoder model. The target autoencoder model is a trained autoencoder model, and the initial autoencoder model is an initially set untrained autoencoder model. Figure 3 It is a schematic diagram of the autoencoder model provided by an embodiment of this application.

[0031] In one embodiment, the first operating parameter includes the parameter data collected in the normal state of the sensor, that is, the parameter data collected in the normal state where the sensor does not fail. The collected data includes but is not limited to the pressure, temperature, flow rate, etc. of the EDU unit. It should be noted that the first operating parameter is a set of parameter data of the sensors, which is composed of the parameter data of one or more sensors. The types of sensors include but are not limited to temperature sensors, pressure sensors, flow sensors, etc.

[0032] In another embodiment, in addition to the parameter data collected in the normal state of the sensor, the first operating parameter also includes data such as the power consumption of the EDU unit and the system efficiency, all of which are used for the autoencoder to reconstruct the input variables of the sensor.

[0033] S1100: Detect the working state of the sensor using the target autoencoder model to obtain the second working state of the sensor.

[0034] In one embodiment, the trained target autoencoder model is used to detect the working state of the real-time running sensor, and it is determined whether the state of the sensor is a fault state. If the second working state of the sensor is a fault state, the deviation value of the sensor is corrected online. If the second working state of the sensor is a non-fault state, that is, the sensor is normal and no fault occurs, the deviation value of the sensor is 0, and the data before and after online correction does not change.

[0035] Figure 4 It is a flowchart for detecting the working state of a sensor through a target autoencoder model provided by an embodiment of the present application, including but not limited to steps S1101 to S1104.

[0036] S1101: Obtain a reference reconstruction deviation value according to the first operation parameter and the reconstructed data output after inputting the first operation parameter into the initial autoencoder model.

[0037] It should be noted that the initial autoencoder model includes an encoder and a decoder, and both the encoder and the decoder include activation functions.

[0038] In one embodiment, a sensor original input variable is established according to the first operation parameter, and the original input variable is used as the input of the activation function of the encoder to obtain an intermediate variable of the original input variable. Then, the intermediate variable of the original input variable is used as the input of the activation function of the decoder to obtain a reconstructed variable of the original input variable, and then a reference reconstruction deviation value is obtained according to the original input variable and the reconstructed variable of the original input variable.

[0039] It can be understood that the sensor original input variable can be the first operation parameter or can be obtained by processing the first operation parameter.

[0040] In one embodiment, the overall obtained by multiplying the original input variable by the weight matrix of the encoder and then adding a bias factor is used as the input of the activation function of the encoder to calculate the intermediate variable of the original input variable. Among them, the activation function of the encoder is a positive saturation linear transfer function. Then, the overall obtained by multiplying the intermediate variable of the original input variable by the weight matrix of the decoder and then adding a bias factor is used as the input of the activation function of the decoder to obtain a reconstructed variable of the original input variable, and then a reference reconstruction deviation value is obtained according to the original input variable and the reconstructed variable of the original input variable. Among them, the activation function of the decoder is a linear transfer function.

[0041] In one embodiment, the first operating parameters are collected multiple times, that is, multiple data sets are collected for a group of sensors. The reconstruction variables of each first operating parameter are obtained by training with an autoencoder model. According to the original input variables and reconstruction variables of each sensor in the multiple data sets, a reference reconstruction deviation value is calculated. It can be understood that the reference reconstruction deviation value is the reconstruction variable deviation threshold of the parameter data collected under the normal state of the sensor, representing the reconstruction deviation range of the normal state data of the sensor.

[0042] S1102: Obtain a working reconstruction deviation value according to the third operating parameter and the target autoencoder model, where the third operating parameter at least includes the operating parameter of the real-time operating state of the sensor.

[0043] In one embodiment, the third operating parameter is the operating parameter collected under the real-time operating state of the unit sensor. The third operating parameter includes the parameter data of a group of sensors and is composed of the parameter data of one or more sensors. Among them, the types of sensors include but are not limited to temperature sensors, pressure sensors, flow sensors, etc.

[0044] In another embodiment, in addition to including the parameter data of the real-time operating state of a group of sensors, the third operating parameter also includes data such as the power consumption of the EDU unit and the system efficiency, all of which are used for the autoencoder to reconstruct the input variables of the sensor.

[0045] It can be understood that if there is one or more sensor failures in a group of sensors, the third operating parameter includes the operating parameter of the sensor failure state; if all sensors in a group are non-faulty sensors, the third operating parameter is the operating parameter of the non-faulty state of the sensor.

[0046] In one embodiment, the third operating parameter is used as the input of the target autoencoder model, and the output obtained is the reconstruction variable of the third operating parameter. Then, according to the difference between the third operating parameter and the reconstruction variable of the third operating parameter, a working reconstruction deviation value is calculated.

[0047] In another embodiment, the third operating parameter is used as the input of the target autoencoder model, and the output obtained is the working reconstruction deviation value, that is, the target autoencoder model includes the calculation of the reconstruction variable and the working reconstruction deviation value.

[0048] It can be understood that the working reconstruction deviation value obtained in step S1102 is the deviation before and after the reconstruction of the real-time operating state parameters of the sensor.

[0049] S1103: Obtain a target fault detection value according to the reference reconstruction deviation value and the working reconstruction deviation value.

[0050] In one embodiment, the ratio of the calculated reference reconstruction deviation value and the working reconstruction deviation value is obtained, and the absolute value and percentage of the ratio are calculated to obtain the target fault detection value.

[0051] S1104: Obtain the second working state of the sensor according to the target fault detection value and the fault detection threshold, where the fault detection threshold is preset.

[0052] In one embodiment, the fault detection threshold is set. The fault detection threshold is set manually, and the threshold unit is in the range of [0, 0.1]. Then, based on the comparison between the target fault detection value and the fault detection threshold, the second working state of the sensor is judged. When the calculated target fault detection value is less than or equal to the fault deviation threshold, it is judged that the sensor has a fault, that is, the second working state of the sensor is the fault state; when the calculated target fault detection value is greater than the fault deviation threshold, it is judged that the sensor has no fault, that is, the second working state of the sensor is the normal state.

[0053] It can be understood that the input of the target autoencoder is a set of parameter data of sensors, and the output of the target autoencoder includes the reconstruction variables or reconstruction variable deviations of each sensor in a set of sensors. The target detection fault values of each sensor can be calculated, and then the working states of each sensor in a set of sensors can be obtained.

[0054] S1200: Obtain the first objective function of the sensor fault deviation according to the second working parameter and the initial autoencoder model, where the second working parameter includes the working parameters of the sensor in the second working state.

[0055] In one embodiment, if the second working state is the sensor fault state, the second working parameter includes the parameter data collected in the sensor fault state, including but not limited to parameter data such as unit temperature, pressure, and flow rate. In another embodiment, if the second working state is the non-fault state of the sensor, the second working parameter includes the parameter data collected in the normal state of the sensor.

[0056] It should be noted that the second working parameter is collected from a set of sensors, where a set of sensors includes one or more sensors. At least one sensor in a set of sensors is in the fault state, or all sensors in a set of sensors are in the normal state. If all sensors in a set of sensors are in the normal state, there is no change in the sensor before and after online calibration.

[0057] In another embodiment, in addition to the parameter data of the sensor in the second working state, the second working parameter further includes data such as the power consumption of the EDU unit and the system efficiency, all of which are used for the autoencoder to reconstruct the input variables.

[0058] It can be understood that both the second operation parameter and the third operation parameter are parameter data collected during the real-time operation of the sensor. The second operation parameter can directly adopt the data of the third operation parameter, that is, the operation state of the sensor is detected through the third operation parameter, and the third operation parameter is corrected; the second operation parameter can also be the parameter data of the sensor collected again during the real-time operation state, that is, the operation state of the sensor is obtained through the third operation parameter, and the parameters in this operation state are collected again for correction.

[0059] In one embodiment, a target fault deviation parameter is set, and then a fourth operation parameter is obtained by adding the second operation parameter and the target fault deviation parameter. The fourth operation parameter is used as the original input variable of the initial autoencoder model. The encoder of the autoencoder model is used to obtain the intermediate variable of the original input variable, and then the intermediate variable is used as the input of the decoder of the initial autoencoder model to obtain the reconstructed variable of the original input variable. A target function for sensor fault deviation is constructed based on the original input variable and the reconstructed variable of the original input variable. Among them, the initial autoencoder model includes an encoder and a decoder. The encoder is used to reduce the dimension of the features of the original input variable to a lower dimension, and the decoder is used to reconstruct the features of the original input variable according to the reduced low-dimensional features.

[0060] Exemplarily, the target fault deviation parameters of each sensor are all set unknown variables. The sum of the second operation parameter and the target fault deviation parameter (i.e., the fourth operation parameter) is the correction function of the sensor. The correction function is used as the original input variable of the initial autoencoder model, and the reference function of the sensor is obtained after being processed by the initial autoencoder model. It can be understood that the reference function of the sensor is the reconstructed variable of the correction function. The square of the difference between the correction function and the reference function is used as the constructed target function.

[0061] S1300: Establish the first prior distribution probability density function and the multivariate normal distribution probability density function of the sensor.

[0062] In one embodiment, the first prior distribution probability density function is set manually according to experience and satisfies a normal distribution with a mean of zero. In addition, a probability density function that satisfies a multivariate normal distribution is set in advance, that is, a multivariate normal distribution probability density function is constructed.

[0063] S1400: Obtain the first posterior distribution probability density function and the first sampling parameter sample according to the first target function, the first prior distribution probability density function, and the multivariate normal distribution probability density function.

[0064] In one embodiment, the first objective function is embedded in the first likelihood function of the sensor fault deviation, that is, the first objective function is used as the independent variable of the first likelihood function of the sensor fault deviation to construct the first likelihood function of the sensor fault deviation. Then, based on the Bayesian inference formula, the first posterior distribution probability density function of the sensor fault deviation is constructed according to the first prior distribution probability density function, the first likelihood function and the total probability density function of the sensor fault deviation, and sampling is performed based on the first posterior distribution probability density function.

[0065] It should be noted that the total probability density function of the sensor fault deviation is a normalization constant.

[0066] In one embodiment, the mean of the first prior distribution probability density function is taken as the initial parameter sample of the Markov chain, and a random number is added to the initial parameter sample to generate the first candidate parameter sample, where the random number is generated from a predefined numerical range. Then, according to the multivariate normal distribution probability density function centered on the initial parameter sample, the multivariate normal distribution probability density function centered on the first candidate parameter sample, the posterior distribution probability density function of the initial parameter sample, and the posterior distribution probability density function of the first candidate parameter sample, the acceptance ratio of the first candidate parameter sample is calculated, and the new parameter sample is judged through the acceptance ratio.

[0067] It should be noted that when the acceptance ratio is greater than the value randomly generated within the range of [0, 1], the first candidate parameter sample is accepted as the first sampling parameter sample. If the acceptance ratio is less than or equal to the value randomly generated within the range of [0, 1], the previous parameter sample is retained as the first sampling parameter sample.

[0068] S1500: Update the first posterior distribution probability density function according to the first sampling parameter sample to obtain the second posterior distribution probability density function.

[0069] In one embodiment, the value of the first sampling parameter sample is added to the second operating parameter as the input of the initial autoencoder model, the calibration function of the sensor is updated, and further the first objective function and the first likelihood function of the sensor fault deviation are updated. In addition, the first prior distribution probability density function is updated according to the first sampling parameter sample. And based on the updated first likelihood function and the updated first prior distribution probability density function, the first posterior distribution probability density function is updated to obtain the second posterior distribution probability density function.

[0070] It can be understood that in this embodiment, the first sampling parameter sample is used as the training sample to update multiple functions such as the objective function, the first prior distribution probability density function, the first posterior distribution probability density function, and the multivariate normal probability density function.

[0071] S1600: Obtain the second sampling parameter sample according to the second posterior distribution probability density function and the multivariate normal probability density function.

[0072] In one embodiment, a second candidate parameter sample is generated by adding a random number to the first sampling parameter sample, where the random number is generated from a predefined numerical range. Calculate the acceptance ratio of the second candidate parameter sample according to the multivariate normal distribution probability density function centered on the first sampling parameter sample, the multivariate normal distribution probability density function centered on the second candidate parameter sample, the posterior distribution probability density function of the first sampling parameter sample, and the posterior distribution probability density function of the second candidate parameter sample.

[0073] In one embodiment, the new parameter sample is judged by the acceptance ratio. When the acceptance ratio is greater than the value randomly generated within the range of [0, 1], the second candidate parameter sample is accepted as the second sampling parameter sample. If the acceptance ratio is less than or equal to the value randomly generated within the range of [0, 1], the first sampling parameter sample is retained as the second sampling parameter sample.

[0074] It should be noted that this embodiment shows the case where the number of iterations is two, and a total of two sampling parameter samples are obtained, namely the first sampling parameter sample and the second sampling parameter sample. It can be understood that the number of iterations can be preset. Preferably, the total number of iterations can be set to no less than 10,000 times. As Figure 5 shown, after sampling in step S2400 each time, in step S2500, update the objective function, prior distribution probability density function, multivariate normal probability density function, posterior distribution probability density function, etc. according to the sampled parameter samples, and sample again according to the updated functions until the number of iterations is reached. The number of new sampled parameter samples obtained is equal to the preset number of iterations, then save all the sampled parameter samples and establish a sample database.

[0075] S1700: Calibrate the second operating parameter according to the first sampling parameter sample and the second sampling parameter sample.

[0076] In one embodiment, obtain the target fault deviation value according to the first sampling parameter sample and the second sampling parameter sample. Calibrate the second operating parameter according to the target fault deviation value. Exemplarily, calculate the mean value of the first sampling parameter sample and the second sampling parameter sample as the target fault deviation value. The sum of the second operating parameter and the target fault deviation value is used as the true measured value of the sensor and fed back to the unit monitoring system for controlling the operation of the EDU circulating water pump and the electric valve. Among them, the target fault deviation value is a value close to the true fault deviation of the sensor.

[0077] It should be noted that in other embodiments, the total number of iterations is set to be greater than 2, and more than two sampling parameter samples are obtained. The mean value of all the sampling parameter samples is also calculated as the target fault deviation value. Preferably, the total number of iterations can be set to no less than 10,000 times.

[0078] In another embodiment, the median value of all the sampling parameter samples is taken as the target fault deviation value. The sum of the second operating parameter and the target fault deviation value is used as the true measured value of the sensor and fed back to the unit monitoring system for controlling and adjusting the operation of the water pump and the electric valve in the EDU. It can be understood that obtaining the target fault deviation value by processing all the sampling parameter samples is within the protection scope of the present application, and the processing includes but is not limited to mean value processing, median value processing, etc.

[0079] It can be understood that after the sensor fails, the collected data has deviations. Since the mean value of all the sampling parameter samples obtained by sampling is closest to the fault deviation value, the second operating parameter with deviations collected is added to the fault deviation value to obtain the closest true measured value of the sensor, which is used to be fed back to the unit. The second operating parameter plus the fault deviation value is the corrected second operating parameter. Among them, the true measured value of the sensor refers to the collected value when the sensor is in a non-fault state under the same conditions.

[0080] Example 1

[0081] Figure 6 It is a flow chart for detecting the operating state of a sensor provided by an example of the present application. As shown in the figure, it includes but is not limited to steps S101 to S106.

[0082] Step 101: Collect the normal state parameters of the unit and respectively establish the original input variable d1 of the sensor.

[0083] Collect a set of parameters such as pressure, temperature, and flow rate of the sensor in a non-fault state, and respectively establish the original input variable d1 of each target sensor according to the collected parameters. Among them, a set of sensors includes one or more sensors, and each of the sensors in a set is a target sensor. In this example, the original input variable d1 is the actually collected measured value of the sensor, and the specific expression is: d1 = f(M) = M, where M is the actually measured value of each target sensor in the normal state.

[0084] Step 102: Obtain the latent variable Z of the original input variable according to the activation function s1 of the encoder.

[0085] It should be noted that the autoencoder model includes an encoder and a decoder. The original input variables are used as the input of the encoder to obtain the latent variable Z of the original input variables. Exemplarily, Z = s1(W1d1 + b1), where W1 is the weight matrix of the encoder, b1 is the bias factor of the encoder, and s1 is the activation function of the encoder. Specifically, s1 is a positive saturated linear transfer function.

[0086] Step 103: Obtain the reconstructed variable d2 of the original input variables according to the activation function s2 of the decoder.

[0087] The latent variable Z of the original input variables obtained in step 102 is used as the input of the decoder to obtain the reconstructed variable d2 of the original input variables. Exemplarily, d2 = s2(W2Z + b2), where W2 is the weight matrix of the decoder, b2 is the bias factor of the decoder, and s2 is the activation function of the decoder. Specifically, s2 is a linear transfer function.

[0088] Step 104: Calculate the reconstructed variable deviation threshold C according to the original input variable d1 and the reconstructed variable d2. s .

[0089] It can be understood that multiple data sets are collected by a group of sensors for training the autoencoder model and calculating the reconstructed variable deviation threshold. Specifically, the calculation formula of the reconstructed variable deviation threshold can be:

[0090]

[0091] where N is the total number of collected data sets, n is the serial number of the collected data sets, I is the total number of a group of sensors, i is the serial number of each target sensor, θ L2 represents L2 regularization, θ sparse represents sparse regularization, λ represents the L2 regularization term coefficient, and β represents the sparse regularization term coefficient.

[0092] Based on multiple data sets in the normal state (i.e., non-fault state) of the sensors, the autoencoder model is trained and the reconstructed deviation threshold is calculated. The obtained reconstructed variable deviation threshold characterizes the reconstructed deviation range in the normal state of the sensors. The reconstructed deviation in the fault state can be compared with it to determine whether the operating state of each target sensor is a fault state.

[0093] Step 105: Collect the real-time operating parameters of the unit and input them into the trained autoencoder model to calculate the reconstructed variable deviation C.

[0094] A set of sensor parameters for collecting the real-time operating state of the unit. It can be understood that all sensors in this set may be in a non-fault state, or one or more of them may be in a fault state. Taking the sensor parameters collected under the real-time operating state of the unit as input, inputting them into the trained autoencoder model to obtain the reconstructed variables of each target sensor, and calculating the reconstructed variable deviation C of each target sensor based on the real-time operating parameters and reconstructed variables of each target sensor.

[0095] Step 106: Set the fault detection threshold E s And calculate the target fault detection value E, and judge the working state of the sensor according to the fault detection threshold E s And the target fault detection value E to judge the working state of the sensor.

[0096] In this example, set the fault detection threshold E s To a value in the range of [0, 0.1] for judging whether there is a fault in the sensor. Specifically, calculate the target fault detection value of each target sensor in a group of sensors and compare them with the fault detection threshold Es respectively. When the target fault detection value of the target sensor is less than the fault detection threshold or the target fault detection value of the target sensor is equal to the fault detection threshold, the target sensor has a fault; when the target fault detection value of the target sensor is greater than the fault detection threshold, the target sensor has no fault. Exemplarily, the calculation method of the target fault detection value is:

[0097]

[0098] Calculate whether the working state of each target sensor in a group of sensors is a fault state according to the above method.

[0099] Example 2

[0100] Figure 7 It is a flowchart for correcting the working state of the sensor provided by an example of the present application. As shown in the figure, it includes but is not limited to steps S201 to S206.

[0101] Step S201: Collect the operating parameters of the real-time operating state of the unit and establish the sensor correction function Y C (x).

[0102] Collect the operating parameters in the fault state of a group of sensors, including but not limited to parameter data such as unit pressure, temperature and flow rate. It can be understood that a group of sensors correspondingly includes multiple working states, and the working states of one or more target sensors are in a fault state.

[0103] Set the sensor fault deviation variable x for each target sensor and establish the correction function Y for each target sensor respectively C(x), where the variable x represents the fault deviation between the data collected in real time by the target sensor and the correct data. Exemplarily, Y C The calculation method of Y

[0104] Y C (x) = f(x, M) = x + M

[0105] Step S202: Take the sensor correction function Y C (x) as the original input variable and input it into the autoencoder model to obtain the sensor reference function Y C (x).

[0106] Take each sensor correction function as the original input variable and input it into the encoder of the autoencoder model to obtain the intermediate variable Z(x) of the original input variable. Take the intermediate variable Z(x) as the input of the decoder of the autoencoder model to obtain the reconstructed variable of each sensor correction function, that is, the sensor reference function Y C (x). Exemplarily, the calculation method of the reconstructed variable of the sensor correction function is:

[0107] Z(x) = s1(W1Y C (x) + b1)

[0108] Y C (x) = s2(W2Z(x) + b2)

[0109] Step S203: Establish the objective function D(x) of the sensor fault deviation according to the sensor correction function Y C (x) and the sensor reference function Y C (x).

[0110] Taking the sensor fault deviation x as the independent variable, construct the objective function D(x) of the sensor fault deviation x of each target sensor according to the sensor correction function and the sensor reference function of each target sensor. Exemplarily, the calculation method of the objective function is:

[0111]

[0112] where N is the total number of the collected data sets, that is, for a group of sensors, a total of N data sets are collected, n is the serial number of the collected data set, I is the total number of a group of sensors, and i is the serial number of each target sensor.

[0113] Step S204: Establish the prior distribution function π(x) of the sensor fault deviation, construct the likelihood function P(Y|x) according to the objective function, and obtain the posterior distribution function P(x|Y) of the sensor fault deviation according to the prior distribution function π(x), the likelihood function P(Y|x) and the total probability function P(Y).

[0114] Establish a prior distribution probability density function π(x) of the sensor fault deviation x for each sensor. It can be understood that due to the complexity of the cooling system, it is difficult to evaluate the fault deviation values with a high probability through human experience. Therefore, set the mean of the prior distribution probability density function π(x) of the sensor fault deviation x to 0, that is, artificially set the prior distribution probability density function π(x) of the sensor fault deviation x to satisfy a normal distribution with a mean of zero.

[0115] Insert the objective function D(x) constructed in step S203 into the likelihood function P(Y|x) of Bayesian inference to achieve the coupling of the autoencoder and Bayesian inference. Based on the calculation formula of Bayesian inference, calculate the posterior distribution probability density function P(x|Y) of the sensor fault deviation x according to the prior distribution probability density function and the likelihood function of the sensor fault deviation x. Specifically, the calculation method of the likelihood function P(Y|x) is as follows:

[0116]

[0117] σ in the above formula is the standard deviation. The calculation method of the posterior distribution probability density function P(x|Y) of the sensor fault deviation x is as follows:

[0118]

[0119] Among them, P(Y) is the total probability density function of the sensor fault deviation x, which is a normalization constant. The calculation method of the total probability density function P(Y) of the sensor fault deviation x is as follows:

[0120] P(Y) = ∫P(Y|x) × π(x)dx

[0121] It can be understood that an objective function is constructed for the sensor fault deviation of each target sensor and inserted into the likelihood function, a prior probability density function of the sensor fault deviation is constructed, and the posterior distribution probability density function of each target sensor is obtained based on Bayesian inference.

[0122] Step S205: Sample the sensor fault deviation through the MCMC algorithm, iteratively update the sensor correction function D(x), the prior distribution probability density function π(x) of the sensor fault deviation, and the posterior distribution probability density function P(x|Y) of the sensor fault deviation according to the sampled data, and save the sampled data to establish a sample database.

[0123] Since it is difficult to directly solve the posterior distribution probability density function \(P(x|Y)\) of the sensor fault deviation \(x\), the Markov Chain Monte Carlo (MCMC) algorithm is used for equivalent sampling. The purpose is to construct a sample database to represent the sensor fault deviation by the mean of the sample database. Specifically, Figure 8 FIG. Figure 8 is a schematic flow chart of MCMC equivalent sampling and iteration provided by an example of the present application. As shown in the figure, the process includes but is not limited to the following steps S2051 to S2056.

[0124] Step S2051: Set the total number of iterations and the initial parameter sample \(X_1\) of the Markov chain.

[0125] Before performing MCMC sampling, the total number of sampling iterations is set manually according to experience. In each iteration, a new parameter sample is generated. Therefore, the total number of samples is equal to the total number of iterations plus one. Preferably, the total number of iterations is not less than 10,000 times.

[0126] Set the prior distribution mean of the sensor fault deviation of each target sensor as the initial parameter sample \(X_1\) for sampling. It can be understood that the initial parameter sample includes the initial parameter samples of each target sensor.

[0127] Step S2052: Set the multivariate normal distribution probability density function \(f(X t * |X t-1 ) and generate a candidate parameter sample \(X t * .

[0128] It should be noted that the multivariate normal probability density function changes with the parameter samples of iterative sampling. Its center is the previous parameter sample \(X t-1 . Its covariance matrix uses a diagonal matrix, and the diagonal of the diagonal matrix is successively the prior distribution standard deviation of the sensor fault deviation \(x\) of each target sensor. \(t\) represents the number of iterations, and \(X t * represents the candidate parameter sample of the \(t\) -th iteration.

[0129] The candidate parameter sample \(X t * is realized by adding a random number within a predefined numerical range to the previous parameter sample \(X t-1 , and it is further determined whether to be accepted based on the multivariate normal distribution probability density function.

[0130] It is understandable that in step S2051, the initial parameter sample X1 of the Markov chain is set, a random number within a predefined numerical range is added to X1 to obtain the candidate parameter sample X2, and then it is determined whether the candidate parameter sample is accepted as the second parameter sample of the Markov chain. Then, the iteration continues until the preset total number of iterations is satisfied.

[0131] Step S2053: Calculate the acceptance ratio of each candidate parameter sample.

[0132] In this example, calculate the acceptance ratio of the candidate parameter samples generated in each iteration of the total number of iterations. Among them, the calculation method of the acceptance ratio of the candidate parameter sample is:

[0133]

[0134] Among them, P(X t * |Y) is the posterior distribution probability density function of the candidate parameter sample X t * , P(X t-1 |Y) is the posterior distribution probability density function of the previous parameter sample X t-1 . f(X t-1 |X t * ) represents the probability density function of the multivariate normal distribution centered on X t * , and f(X t * |X t-1 ) represents the probability density function of the multivariate normal distribution centered on X t-1 .

[0135] Step S2054: Select a new parameter sample according to the acceptance ratio.

[0136] In this step, in each iteration, compare the calculated acceptance ratio α of the candidate parameter sample with the value randomly generated in [0, 1]. When the acceptance ratio α is greater than the randomly generated value, accept the candidate parameter sample X t * as the new parameter sample. If the acceptance ratio α is less than or equal to the randomly generated value, retain the previous parameter sample X t-1 as the new parameter sample.

[0137] Step S2055: Update the sensor correction function D(x), the prior distribution probability density function π(x) of the sensor fault deviation, and the posterior distribution probability density function P(x|Y) of the sensor fault deviation according to the new parameter samples generated in each iteration.

[0138] In this example, the newly selected parameter samples at each iteration are the sampling data. The sampling data is used to update the calibration functions of each target sensor. The updated calibration functions serve as the new training sample data for the autoencoder, obtaining the new reference functions of each target sensor. Based on the updated calibration functions and reference functions of each target sensor, the sensor fault deviation objective function of each target sensor is updated.

[0139] Then, the updated objective function is inserted into the likelihood function. At the same time, the sampling data is used as the sample of the prior distribution probability density function to update the prior distribution probability density function. Finally, based on the updated prior distribution probability density function and the updated likelihood function, the posterior distribution probability density function is updated.

[0140] At the next sampling, the acceptance ratio of the candidate parameter samples is calculated according to the posterior distribution probability density function of each target sensor updated in the previous time, so as to further sample.

[0141] This application couples the autoencoder and Bayesian inference to achieve sample sampling. Specifically, the sensor fault deviation values obtained by Bayesian inference sampling are used to train the autoencoder, and the objective function, prior distribution probability density function, posterior distribution density probability function, etc. are iteratively updated. The updated posterior distribution density probability function further affects the new sample sampling. This application uses embedded coupled sampling to obtain the optimal value of the sensor fault deviation, abandons the conventional method of establishing parameter associations through complex physical definitions, does not need to know and establish an accurate physical mechanism model, sets unknown variables and iteratively updates the sampling to obtain the optimal value of the unknown variables, realizes the online calibration of sensor faults, and improves the service life of the sensor.

[0142] Step S2056: Repeat sampling and update iteration until the set sampling quantity is reached to obtain all parameter samples.

[0143] In step S2051, the total number of iterations is set. Repeat steps S2052 to S2055 for sampling and updating until the set total number of iterations is reached. The initial parameter samples and all newly sampled parameter samples form a sample database.

[0144] Step S2057: Statistically analyze all samples to obtain parameters such as the mean and median of the samples.

[0145] Statistically analyze all samples in the sample database to obtain parameters such as the mean and median of the samples, which are used to determine the sensor fault deviation value. It can be understood that for each target sensor in a group of sensors, the mean, median, etc. parameters are sampled and obtained through the embedded coupling method of the autoencoder and Bayesian inference, and are used to calculate the sensor fault deviation of each target sensor.

[0146] Step S206: Statistically analyze the sample data, take the mean value of the sample data in the sample database as the sensor fault deviation value, and perform online calibration of the sensor according to the sensor fault deviation value.

[0147] It can be understood that the posterior distribution probability density function satisfies a normal distribution. When the sampled data is relatively large, both the sample mean and the sample median are close to the true fault deviation value of the sensor. When the sampled data is relatively small, the sample mean is closer to the true fault deviation value of the sensor. Therefore, preferably, the mean value of the sample data in the sample database is taken as the sensor fault deviation value. However, selecting parameters such as the median of the sample data as the sensor fault deviation value is also within the scope of protection of this application.

[0148] After determining the sample data mean value, use the sample data mean value as the sensor fault deviation value and feedback it to the unit monitoring system to achieve online calibration of the real-time measurement data of the sensor.

[0149] It should be noted that in this application, a group of sensors is used as the input of the autoencoder, the objective function of the sensor fault deviation of each target sensor is constructed, and the posterior distribution probability density function of each target sensor is calculated. Further sampling and calculation of the sample mean value are performed to obtain the sensor fault deviation value of each target sensor.

[0150] Taking a group of sensors including 5 sensors, the types of which are not limited to temperature, pressure, and flow sensors as an example, in step S206, the obtained sensor fault deviation value is [0, 0.6, 0, 0, 0], and the real-time measurement data of this group of sensors is [100, 32, 200, 280, 50]. According to the sensor fault deviation values, it can be known that the second sensor is in a fault state, and the other four sensors are in a normal working state. The true values of each target sensor after online calibration are [100, 32.6, 200, 280, 50]. In step S206, the obtained sensor fault deviation value is [0, -0.6, 0, 0, 0], and the real-time measurement data of this group of sensors is [100, 32, 200, 280, 50]. According to the sensor fault deviation values, it can be known that the second sensor is in a fault state, and the other four sensors are in a normal working state. The true values of each target sensor after online calibration are [100, 31.4, 200, 280, 50]

[0151] In step S206, the obtained sensor fault deviation value is [0, 0, 0, 0, 0], and the real-time measurement data of this group of sensors is [100, 32, 200, 280, 50]. According to the sensor fault deviation values, it can be known that all sensors are in a normal working state. The true values of each target sensor after online calibration are [100, 32, 200, 280, 50].

[0152] In step S206, the sensor fault deviation value is obtained as [0, -0.6, 0, 1, 0], and the real-time measurement data of this group of sensors is [100, 32, 200, 280, 50]. According to the sensor fault deviation values of each sensor, it can be known that the second sensor and the fourth sensor are in a fault state, and the other three sensors are in a normal working state. The true values of each target sensor after online calibration are [100, 31.4, 200, 281, 50].

[0153] Example 3

[0154] Figure 9 It is a flowchart of sensor operation status detection and calibration provided by an example of the present application. As shown in the figure, it includes but is not limited to steps S301 to S312.

[0155] Step 301: Collect the normal state parameters of the unit and establish the original input variable d1 of the sensor respectively.

[0156] Collect parameters such as pressure, temperature, and flow rate of a group of sensors in a non-fault state, and establish the original input variable d1 of each target sensor according to the collected parameters. Among them, a group of sensors includes one or more sensors, and each in a group of sensors is a target sensor. In this example, the original input variable d1 is the actually collected sensor measurement value, and the specific expression is: d1 = f(M) = M, where M is the actual measurement value of each target sensor in the normal state.

[0157] Step 302: Obtain the latent variable Z of the original input variable according to the activation function s1 of the encoder.

[0158] It should be noted that the autoencoder model includes an encoder and a decoder. Taking the original input variable as the input of the encoder, the latent variable Z of the original input variable is obtained. Exemplarily, Z = s1(W1d1 + b1), where W1 is the weight matrix of the encoder, b1 is the bias factor of the encoder, and s1 is the activation function of the encoder. Specifically, s1 is a positive saturation linear transfer function.

[0159] Step 303: Obtain the reconstructed variable d2 of the original input variable according to the activation function s2 of the decoder.

[0160] Taking the latent variable Z of the original input variable obtained in step 102 as the input of the decoder, the reconstructed variable d2 of the original input variable is obtained. Exemplarily, d2 = s2(W2Z + b2), where W2 is the weight matrix of the decoder, b2 is the bias factor of the decoder, and s2 is the activation function of the decoder. Specifically, s2 is a linear transfer function.

[0161] Step 304: Calculate the reconstruction variable deviation threshold C based on the original input variable d1 and the reconstructed variable d2 s .

[0162] It can be understood that for a group of sensors, multiple data sets are collected for training the autoencoder model and calculating the reconstruction variable deviation threshold. Specifically, the calculation formula for the reconstruction variable deviation threshold can be:

[0163]

[0164] where N is the total number of collected data sets, n is the serial number of the collected data sets, I is the total number of a group of sensors, i is the serial number of each target sensor, θ L2 represents L2 regularization, θ sparse represents sparse regularization, λ represents the L2 regularization term coefficient, and β represents the sparse regularization term coefficient.

[0165] Based on multiple data sets in the normal state (i.e., non-fault state) of the sensors, the autoencoder model is trained and the reconstruction deviation threshold is calculated. The obtained reconstruction variable deviation threshold characterizes the reconstruction deviation range in the normal state of the sensors. The reconstruction deviation in the fault state can be compared with it to determine whether the operating state of each target sensor is a fault state.

[0166] Step 305: Collect the operating parameters of the real-time operating state of the unit and input them into the trained autoencoder model to calculate the reconstruction variable deviation C.

[0167] Collect a group of sensor parameters of the real-time operating state of the unit. It can be understood that all sensors in this group may be in a non-fault state, or there may be one or more in a fault state. Take the sensor parameters collected under the real-time operating state of the unit as the input and input them into the trained autoencoder model to obtain the reconstruction variable of each target sensor, and calculate the reconstruction variable deviation C of each target sensor according to the real-time operating parameters and the reconstruction variable of each target sensor.

[0168] Step 306: Set the fault detection threshold E s and calculate the target fault detection value E. Obtain the sensor operating state according to the fault detection threshold E s and the target fault detection value E.

[0169] In this example, set the fault detection threshold E s as a value in the threshold range of [0, 0.1] for judging whether there is a fault in the sensor. Specifically, calculate the target fault detection value of each target sensor in a group of sensors and compare them with the fault detection threshold E sCompare. When the target fault detection value of the target sensor is less than the fault detection threshold, or when the target fault detection value of the target sensor is equal to the fault detection threshold, the target sensor is in a fault state; when the target fault detection value of the target sensor is greater than the fault detection threshold, the target sensor is in a normal state. Exemplarily, the calculation method of the target fault detection value is as follows:

[0170]

[0171] Step S307: Collect the operation parameters of the unit in the sensor operation state obtained in S306, and establish a sensor calibration function Y C (x).

[0172] Collect the operation parameters of this group of sensors in the real-time operation state, including but not limited to parameters such as unit pressure, temperature, and flow rate. It can be understood that a group of sensors corresponds to multiple working states.

[0173] Set a sensor fault deviation variable x for each target sensor, and establish a calibration function Y C (x) for each target sensor respectively, where the variable x represents the fault deviation between the data collected by the target sensor in real time and the correct data. Exemplarily, the calculation method of Y C (x) is as follows:

[0174] Y C (x) = f(x, M) = x + M

[0175] Step S308: Input the sensor calibration function Y C (x) as the original input variable into the autoencoder model to obtain the sensor reference function Y C (x).

[0176] Input each sensor calibration function as the original input variable into the encoder of the autoencoder model to obtain the intermediate variable Z(x) of the original input variable. Input the intermediate variable Z(x) into the decoder of the autoencoder model to obtain the reconstructed variable of each sensor calibration function, that is, the sensor reference function Y C (x). Exemplarily, the calculation method of the reconstructed variable of the sensor calibration function is as follows:

[0177] Z(x) = s1(W1Y C (x) + b1)

[0178] Y C (x) = s2(W2Z(x) + b2)

[0179] Step S309: According to the sensor calibration function Y C (x) and the sensor reference function YC (x) Establish the objective function D(x) of the sensor fault deviation.

[0180] Taking the sensor fault deviation x as the independent variable, construct the objective function D(x) of the sensor fault deviation x for each target sensor according to the sensor calibration function and the sensor reference function of each target sensor. Exemplarily, the calculation method of the objective function is:

[0181]

[0182] Where N is the total number of the collected data sets, that is, for a group of sensors, a total of N data sets are collected, n is the serial number of the collected data set, I is the total number of a group of sensors, and i is the serial number of each target sensor.

[0183] Step S310: Establish the prior distribution function π(x) of the sensor fault deviation, construct the likelihood function P(Y|x) according to the objective function, and obtain the posterior distribution function P(x|Y) of the sensor fault deviation according to the prior distribution function π(x) and the likelihood function P(Y|x).

[0184] Establish the prior distribution probability density function π(x) of the sensor fault deviation x for each sensor. It can be understood that due to the complexity of the cooling system, it is usually difficult to evaluate the fault deviation values with greater possibility through human experience. Therefore, set the mean value of the prior distribution probability density function π(x) of the sensor fault deviation x to 0, that is, artificially set the prior distribution probability density function π(x) of the sensor fault deviation x to satisfy the normal distribution with a mean of zero.

[0185] Insert the objective function D(x) constructed in step S309 into the likelihood function P(Y|x) of Bayesian inference to realize the coupling of the autoencoder and Bayesian inference. Based on the calculation formula of Bayesian inference, calculate the posterior distribution probability density function P(x|Y) of the sensor fault deviation x according to the prior distribution probability density function and the likelihood function of the sensor fault deviation x. Specifically, the calculation method of the likelihood function P(Y|x) is:

[0186]

[0187] σ in the above formula is the standard deviation. The calculation method of the posterior distribution probability density function P(x|Y) of the sensor fault deviation x is:

[0188]

[0189] Where P(Y) is the total probability density function of the sensor fault deviation x, which is a normalization constant. The calculation method of the total probability density function P(Y) of the sensor fault deviation x is:

[0190] P(Y) = ∫P(Y|x)×π(x)dx

[0191] It can be understood that an objective function is constructed for the sensor fault deviation of each target sensor and inserted into the likelihood function to construct the prior probability density function of the sensor fault deviation, and the posterior distribution probability density function of each target sensor is obtained based on Bayesian inference.

[0192] Step S311: Sample the sensor fault deviation through the MCMC algorithm, iteratively update the sensor correction function D(x), the prior distribution probability density function π(x) of the sensor fault deviation, and the posterior distribution probability density function P(x|Y) of the sensor fault deviation according to the sampled data, save the sampled data, and establish a sample database.

[0193] Since it is difficult to directly solve the posterior distribution probability density function P(x|Y) of the sensor fault deviation x, equivalent sampling is performed through the Markov Chain Monte Carlo (MCMC) algorithm. The purpose is to construct a sample database to represent the sensor fault deviation by the mean value of the sample database.

[0194] Specifically, referring to the MCMC equivalent sampling and iterative process in Example 2, this example will not be elaborated.

[0195] Step S312: Statistically analyze the sample data, take the mean value of the sample data in the sample database as the sensor fault deviation value, and perform online correction of the sensor according to the sensor fault deviation value.

[0196] It can be understood that the posterior distribution probability density function satisfies the normal distribution. When there are more sampled data, both the sample mean and the sample median are close to the true fault deviation value of the sensor. When there are fewer sampled data, the sample mean is closer to the true fault deviation value of the sensor. Therefore, preferably, the mean value of the sample data in the sample database is taken as the sensor fault deviation value, but selecting parameters such as the median of the sample data as the sensor fault deviation value is also within the protection scope of this application.

[0197] After determining the sample data mean value, use the sample data mean value as the sensor fault deviation value and feedback it to the unit monitoring system to achieve online correction of the real-time measurement data of the sensor.

[0198] Figure 10 It is a schematic structural diagram of a computer device provided in an embodiment of the present application, as Figure 10As shown, the computer device includes a processor 3000, a memory 3100, and a computer program stored on the memory 3100 and executable on the processor 3000. The processor 3000 executes the calibration method of the sensor provided in any embodiment of the present application. Among them, the number of the processor 3000 and the memory 3100 can be one or more. In the figure, one processor 3000 and one memory 3100 are taken as examples; the processor 3000 and the memory 3100 in the device can be connected by a bus or other means. In the figure, connection by a bus is taken as an example.

[0199] An embodiment of the present application further provides a computer-readable storage medium storing computer-executable instructions for executing the calibration method of the sensor provided in any embodiment of the present application.

[0200] An embodiment of the present application further provides a computer program product including a computer program or computer instructions. The computer program or computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer program or computer instructions from the computer-readable storage medium, and the processor executes the computer program or computer instructions, so that the computer device executes the calibration method of the sensor provided in any embodiment of the present application.

[0201] The system architecture and application scenarios described in the embodiments of the present application are to more clearly illustrate the technical solutions of the embodiments of the present application, and do not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those skilled in the art can know that with the evolution of the system architecture and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0202] Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, and their appropriate combinations.

[0203] In a hardware implementation, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be executed by the cooperation of several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or may be implemented as hardware, or may be implemented as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or transitory medium). As is well known to those of ordinary skill in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those of ordinary skill in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transmission mechanism, and may include any information delivery medium.

[0204] Some embodiments of the present application have been illustrated above with reference to the accompanying drawings, and thus the scope of the rights of the present application is not limited thereby. Any modifications, equivalent replacements, and improvements made by those skilled in the art without departing from the scope and essence of the present application shall be within the scope of the rights of the present application.

Claims

1. A calibration method for a sensor, characterized in that, The method includes: Obtaining a target autoencoder model according to first operation parameters and an initial autoencoder model, where the first operation parameters at least include operation parameters of the normal state of the sensor; Detecting the working state of the sensor using the target autoencoder model to obtain a second operation state of the sensor; Obtaining a first objective function of the sensor fault deviation according to second operation parameters and the initial autoencoder model, where the second operation parameters include operation parameters of the sensor in the second operation state; Establishing a first prior distribution probability density function and a multivariate normal distribution probability density function of the sensor; Obtaining a first posterior distribution probability density function and a first sampling parameter sample according to the first objective function, the first prior distribution probability density function, and the multivariate normal distribution probability density function; Updating the first posterior distribution probability density function according to the first sampling parameter sample to obtain a second posterior distribution probability density function; Obtaining a second sampling parameter sample according to the second posterior distribution probability density function and the multivariate normal distribution probability density function; Calibrating the second operation parameters according to the first sampling parameter sample and the second sampling parameter sample.

2. The calibration method according to claim 1, characterized in that The step of detecting the working state of the sensor using the target autoencoder model to obtain a second operation state of the sensor includes: Obtaining a reference reconstruction deviation value according to the first operation parameters and the reconstructed data output after inputting the first operation parameters into the initial autoencoder model; Obtaining a working reconstruction deviation value according to third operation parameters and the target autoencoder model, where the third operation parameters at least include operation parameters of the real-time running state of the sensor; Obtaining a target fault detection value according to the reference reconstruction deviation value and the working reconstruction deviation value; Obtaining the second operation state of the sensor according to the target fault detection value and a fault detection threshold, where the fault detection threshold is preset.

3. The calibration method according to claim 1 or 2, characterized in that The step of obtaining a first objective function of the sensor fault deviation according to the second operation parameters and the initial autoencoder model includes: Setting a target fault deviation parameter; Obtaining fourth operation parameters according to the second operation parameters and the target fault deviation parameter; Obtaining a first reconstruction variable according to the fourth operation parameters and the initial autoencoder model; Obtaining a first objective function of the sensor fault deviation according to the fourth operation parameters and the first reconstruction variable.

4. The calibration method according to claim 3, wherein The step of obtaining a first posterior distribution probability density function and a first sampling parameter sample according to the first objective function, the first prior distribution probability density function, and the multivariate normal distribution probability density function includes: Constructing a first likelihood function according to the first objective function; Obtaining a first posterior distribution probability density function of the sensor fault deviation according to the likelihood function, the first prior distribution probability density function, and a total probability density function, where the total probability density function is a normalization constant; Based on the first prior distribution probability density function and the multivariate normal distribution probability density function, a first sampling parameter sample is obtained.

5. The calibration method according to claim 4, characterized in that, The obtaining of the first sampling parameter sample based on the first prior distribution probability density function and the multivariate normal distribution probability density function includes: Based on the first prior distribution probability density function, an initial parameter sample is obtained; Based on the initial parameter sample and a random value, a first candidate parameter sample is obtained; Based on the multivariate normal distribution probability density function, the first posterior distribution probability density function, and the first candidate parameter sample, a first acceptance parameter corresponding to the first candidate parameter sample is calculated; Based on the first acceptance parameter, a first sampling parameter sample is obtained.

6. The calibration method according to claim 5, wherein The updating of the first posterior distribution probability density function based on the first sampling parameter sample to obtain a second posterior distribution probability density function further includes: Based on the first sampling parameter sample, the first objective function and the first likelihood function are updated to obtain a second objective function and a second likelihood function; Based on the first sampling parameter sample, the first prior distribution probability density function is updated to obtain the second prior distribution probability density function.

7. The calibration method according to claim 6, wherein The obtaining of the second sampling parameter sample based on the second posterior distribution probability density function and the multivariate normal distribution probability density function includes: Based on the first sampling parameter sample and a random value, a second candidate parameter sample is obtained; Based on the multivariate normal distribution probability density function, the second posterior distribution probability density function, and the second candidate parameter sample, a second acceptance parameter corresponding to the second candidate parameter sample is calculated; Based on the second acceptance parameter, a second sampling parameter sample is obtained.

8. The calibration method according to claim 7, characterized in that, The correcting of the second operation parameter based on the first sampling parameter sample and the second sampling parameter sample includes: Based on the first sampling parameter sample and the second sampling parameter sample, a target fault deviation value is obtained; Based on the target fault deviation value, the second operation parameter is corrected.

9. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the correction method of the sensor according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium storing computer-executable instructions for executing the correction method of the sensor according to any one of claims 1 to 8.

11. A computer program product storing computer-executable instructions for executing the correction method of the sensor according to any one of claims 1 to 8.

Images