CN-121977747-A - Real-time fault detection method of pressure sensor in dynamic environment

CN121977747ACN 121977747 ACN121977747 ACN 121977747ACN-121977747-A

Abstract

The invention discloses a real-time fault detection method of a pressure sensor in a dynamic environment, which relates to the technical field of data analysis and comprises the steps of obtaining historical reading signal sample data of the pressure sensor, carrying out state trend matching screening on the historical reading signal sample data of the pressure sensor and known fault types of the pressure sensor, obtaining a set of pointing fault types of the historical reading signal sample data of the pressure sensor, establishing a classification model of the known fault types of the pressure sensor, taking the real-time reading signal sample data of the pressure sensor as input, generating a confidence coefficient of the pointing fault types of the real-time reading signal sample data of the pressure sensor, establishing a real-time fault type detection decision tree, generating a real-time fault diagnosis classification of the pressure sensor, establishing a state transition matrix of the pressure sensor, and determining a real-time fault state of the pressure sensor. The method has the beneficial effects that the robustness and the reliability of fault detection in a dynamic environment are remarkably improved.

Inventors

HE JIANCHAO
ZHANG ZHONG
LEI FANGFANG
ZHANG DETAO
LI DAOLU
Pan Pengdong

Assignees

东莞市帝恩检测有限公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (7)

1. The real-time fault detection method of the pressure sensor in the dynamic environment is characterized by comprising the following steps: s1, acquiring historical reading signal sample data of a pressure sensor based on a background database, analyzing the change trend of the pressure sensor in the historical reading signal sample data, generating a reading signal sample trend vector of the pressure sensor, and carrying out state trend matching screening with the known fault types of the pressure sensor to obtain a set of pointing fault types of the historical reading signal sample data of the pressure sensor; S2, according to a set of fault types pointed by historical reading signal sample data of the pressure sensor, a known fault type classification model of the pressure sensor is established, a super plane of optimal classification of the fault types pointed by the historical reading signal sample data is generated, real-time reading signal sample data of the pressure sensor is taken as input, and a confidence degree of the fault types pointed by the real-time reading signal sample data of the pressure sensor is generated; S3, based on the confidence that the real-time reading signal sample data of the pressure sensor points to the fault type, a real-time fault type detection decision tree is established, real-time fault diagnosis classification of the pressure sensor is generated, a state transition matrix of the pressure sensor is established, and the real-time fault state of the pressure sensor is determined.
2. The method for detecting a real-time failure of a pressure sensor in a dynamic environment according to claim 1, wherein step S1 specifically comprises: Preprocessing historical reading signal sample data of the pressure sensor, sliding a window, taking unit time as an observation window, taking the historical reading signal sample data as an observation object, acquiring a time sequence of the historical reading signal sample of the pressure sensor, extracting time domain characteristics of the historical reading signal sample of the pressure sensor, substituting the time domain characteristics into FFT (fast Fourier transform) to obtain a frequency domain signal of the historical reading signal sample of the pressure sensor; according to PSD power spectral density, extracting main frequency, bandwidth and spectrum centroid in a frequency domain signal of a historical reading signal sample of the pressure sensor to obtain the frequency domain characteristic of the historical reading signal sample of the pressure sensor; According to the historical reading signal sample frequency domain characteristics of the pressure sensor, training time sequence autoregressive, taking the past time step value of the historical reading signal sample frequency domain characteristics as input, taking the current time step of the historical reading signal sample frequency domain characteristics as output, and generating the reading signal sample trend vector of the pressure sensor by taking the historical reading signal sample frequency domain characteristics as an error function.
3. The method for detecting a real-time failure of a pressure sensor in a dynamic environment according to claim 2, wherein step S2 further comprises: Based on the read signal sample data of the known fault type of the pressure sensor, extracting the read signal sample time domain characteristics of the known fault type of the pressure sensor through a sliding window, substituting the read signal sample time domain characteristics into FFT (fast Fourier transform), and extracting the read signal sample frequency domain characteristics of the known fault type of the pressure sensor through PSD (phase-sensitive detector) power spectrum density; Based on the frequency domain characteristics of the read signal samples of the known fault type of the pressure sensor, determining the prior probability of the known fault type of the pressure sensor according to Bayesian prior, and observing the probability of the frequency domain characteristics of the read signal samples of the known fault type of the pressure sensor when the known fault type is given according to a likelihood function; Using the prior probability of the known fault type of the pressure sensor and the probability of the frequency domain feature of the read signal sample of the known fault type of the pressure sensor when the known fault type is given, constructing a Bayesian posterior, and obtaining the triggering posterior probability of the known fault type of the pressure sensor when the frequency domain feature of the read signal sample of the known fault type is observed; According to the maximum likelihood estimation, extracting the frequency domain characteristics of the read signal samples of the known fault type from the trigger posterior probability of the known fault type of the pressure sensor, and constructing the frequency domain characteristic distribution vector of the read signal samples of the known fault type of the pressure sensor; based on cosine similarity, calculating similarity between a read signal sample trend vector of the pressure sensor and a read signal sample frequency domain feature distribution vector of a known fault type of the pressure sensor, and screening according to positive similarity between the read signal sample trend vector and the read signal sample frequency domain feature distribution vector of the known fault type to obtain a historical read signal sample data pointing fault type set of the pressure sensor.
4. The method for detecting a real-time failure of a pressure sensor in a dynamic environment according to claim 3, wherein step S2 specifically comprises: Based on the historical reading signal sample data of the pressure sensor pointing to the fault type set and the reading signal sample data of the known fault type of the pressure sensor, the historical reading signal sample data of the pressure sensor pointing to positive and negative samples of the fault type is built; Performing standardized processing on positive and negative samples of the fault type pointed by historical reading signal sample data of the pressure sensor, substituting the positive and negative samples into a PCA principal component analysis method to obtain positive and negative dimension reduction samples of the fault type pointed by the historical reading signal sample data of the pressure sensor; Based on an SVM support vector machine, according to One-vs-Rest classification strategy, a radial basis function is used as a kernel function to construct a fault classification model of the pressure sensor and a non-fault classification model of the pressure sensor, historical reading signal sample data of the pressure sensor points to positive and negative dimension reduction samples of fault types to be used as input, and historical reading signal sample data of the pressure sensor points to an optimal classification hyperplane of fault classification and non-fault classification to be used as output.
5. The method for real-time fault detection of a pressure sensor in a dynamic environment according to claim 4, wherein step S2 further comprises: Based on time-frequency analysis, marking the frequency domain information of the real-time reading signal sample of the pressure sensor, extracting the frequency domain vector of the real-time reading signal sample of the pressure sensor by using PSD power spectral density, and substituting the frequency domain vector into PCA main component analysis for dimension reduction processing; Calculating the space distance between the real-time reading signal sample frequency domain vector of the pressure sensor and the optimal classification hyperplane of the historical reading signal sample data pointing fault classification and non-fault classification of the pressure sensor according to Euclidean distance to obtain the real-time reading signal sample pointing fault classification and non-fault classification optimal classification hyperplane distance vector of the pressure sensor; According to the real-time reading signal sample pointing fault classification and non-fault classification optimal classification hyperplane distance vector of the pressure sensor, generating a real-time reading signal sample data pointing fault type and non-fault type confidence coefficient score of the pressure sensor through a substituted monotonically decreasing function, and assembling the real-time reading signal sample data pointing fault type and non-fault type confidence coefficient vector of the pressure sensor.
6. The method for detecting a real-time failure of a pressure sensor in a dynamic environment according to claim 5, wherein step S3 specifically comprises: Based on random forests, constructing a real-time fault type detection decision tree, taking historical reading signal sample data pointing fault classification and non-fault classification in an optimal classification hyperplane of the pressure sensor as branch nodes according to the historical reading signal sample data pointing fault classification and non-fault classification, taking the confidence of the real-time reading signal sample data pointing fault type of the pressure sensor as leaf nodes, taking the real-time reading signal sample data of the pressure sensor as input, and taking the maximum value of information gain of the real-time reading signal sample data relative to the branch nodes as a segmentation condition to generate real-time fault diagnosis classification of the pressure sensor.
7. The method for real-time fault detection of a pressure sensor in a dynamic environment according to claim 6, wherein step S3 further comprises: Determining known fault types and non-fault reading signal sample data of the pressure sensor, constructing an initial state transition matrix of the fault types of the pressure sensor, taking the known fault types and the non-fault types as parent nodes, taking real-time reading signal sample data pointing to real-time fault diagnosis classification as attribute nodes, marking the trigger frequency of child nodes under each parent node by unit time stamps, calculating transition probability among each parent node in the initial state transition matrix according to Markov chains, and determining the real-time fault state of the pressure sensor.

Description

Real-time fault detection method of pressure sensor in dynamic environment Technical Field The invention relates to the technical field of data analysis, in particular to a real-time fault detection method of a pressure sensor in a dynamic environment. Background The existing pressure sensor dynamic environment real-time fault detection method is difficult to adapt to dynamic noise interference based on a fixed threshold or a detection mechanism with simple statistical characteristics, so that the false alarm rate is high, the characteristic representation capability of a time domain analysis method on non-stationary signals is insufficient, deep fusion of a frequency domain and a trend is lacking, a traditional classification model (such as a single SVM or a neural network) does not effectively integrate real-time confidence and historical state evolution rules, time sequence dependence of fault development is ignored, most schemes depend on a large amount of labeling data and model update is lagged, online self-adaptive learning cannot be realized, and a system lacks interpretable decision paths and state transition tracking, so that the fault root analysis and early warning capability is limited. The existing method has the problems of unbalanced sensitivity and stability, and difficult real-time performance and accuracy in a dynamic environment. Disclosure of Invention In order to solve the technical problems, the technical scheme solves the problems by providing a real-time fault detection method of the pressure sensor in a dynamic environment. In order to achieve the above purpose, the invention adopts the following technical scheme: the real-time fault detection method of the pressure sensor in the dynamic environment comprises the following steps: s1, acquiring historical reading signal sample data of a pressure sensor based on a background database, analyzing the change trend of the pressure sensor in the historical reading signal sample data, generating a reading signal sample trend vector of the pressure sensor, and carrying out state trend matching screening with the known fault types of the pressure sensor to obtain a set of pointing fault types of the historical reading signal sample data of the pressure sensor; S2, according to a set of fault types pointed by historical reading signal sample data of the pressure sensor, a known fault type classification model of the pressure sensor is established, a super plane of optimal classification of the fault types pointed by the historical reading signal sample data is generated, real-time reading signal sample data of the pressure sensor is taken as input, and a confidence degree of the fault types pointed by the real-time reading signal sample data of the pressure sensor is generated; S3, based on the confidence that the real-time reading signal sample data of the pressure sensor points to the fault type, a real-time fault type detection decision tree is established, real-time fault diagnosis classification of the pressure sensor is generated, a state transition matrix of the pressure sensor is established, and the real-time fault state of the pressure sensor is determined. Preferably, step S1 specifically includes: Preprocessing historical reading signal sample data of the pressure sensor, sliding a window, taking unit time as an observation window, taking the historical reading signal sample data as an observation object, acquiring a time sequence of the historical reading signal sample of the pressure sensor, extracting time domain characteristics of the historical reading signal sample of the pressure sensor, substituting the time domain characteristics into FFT (fast Fourier transform) to obtain a frequency domain signal of the historical reading signal sample of the pressure sensor; according to PSD power spectral density, extracting main frequency, bandwidth and spectrum centroid in a frequency domain signal of a historical reading signal sample of the pressure sensor to obtain the frequency domain characteristic of the historical reading signal sample of the pressure sensor; According to the historical reading signal sample frequency domain characteristics of the pressure sensor, training time sequence autoregressive, taking the past time step value of the historical reading signal sample frequency domain characteristics as input, taking the current time step of the historical reading signal sample frequency domain characteristics as output, and generating the reading signal sample trend vector of the pressure sensor by taking the historical reading signal sample frequency domain characteristics as an error function. Preferably, step S2 further includes: Based on the read signal sample data of the known fault type of the pressure sensor, extracting the read signal sample time domain characteristics of the known fault type of the pressure sensor through a sliding window, substituting the read signal sample time domain characteristics into FFT