Intelligent fault diagnosis method and system for multi-channel gas sensor

Abstract

The invention relates to the technical field of fault diagnosis, and discloses an intelligent fault diagnosis method and system for a multi-channel gas sensor, and the method comprises the steps: carrying out the random geometric analysis and response limitation region construction of a multi-channel gas sensor array, and obtaining an effective sensor set; performing correlation analysis and sensitivity precision dilution measurement function screening to obtain an optimal sensor combination; performing periodic sampling and feature extraction to obtain a feature data matrix; performing deep Q network analysis to generate a fault diagnosis trigger signal; performing sensor channel block mapping and position coding on the feature data matrix according to the fault diagnosis trigger signal to obtain a space-time separation feature vector; and inputting the space-time separation feature vector into a bidirectional long-short-term memory network to perform fault type determination, and outputting a fault type determination result, thereby realizing accurate identification and classification of different types of faults of the multi-channel gas sensor.

Description

Technical Field

[0001] The present application relates to the technical field of fault diagnosis, and in particular to an intelligent fault diagnosis method and system for a multi-channel gas sensor. Background Art

[0002] Multi-channel gas sensor systems are often composed of a large number of sensors forming a complex sensor network. These sensors are prone to various faults during long-term operation, such as response offset, reduced sensitivity, cross interference and other problems, which seriously affect the reliability and measurement accuracy of the system.

[0003] Traditional gas sensor fault diagnosis methods mainly rely on single threshold judgment or simple statistical analysis, which is difficult to deal with complex fault scenarios in high-density sensor networks. Especially when there are a large number of sensors and a complex working environment, traditional methods are difficult to effectively identify and handle various faults, and cannot fully utilize the collaborative relationship between sensors to improve the accuracy of fault diagnosis. Summary of the invention

[0004] The present application provides an intelligent fault diagnosis method and system for a multi-channel gas sensor, thereby achieving accurate identification and classification of different types of faults of the multi-channel gas sensor.

[0005] The first aspect of the present application provides an intelligent fault diagnosis method for a multi-channel gas sensor, the intelligent fault diagnosis method for the multi-channel gas sensor comprising: The response data of n sensor channels in the multi-channel gas sensor array are subjected to random geometric analysis and response restriction region construction to obtain an effective sensor set. Performing correlation analysis and sensitivity precision dilution measurement function screening on response data between sensor channels in the effective sensor set to obtain an optimal sensor combination; Performing periodic sampling and feature extraction on the response data of the optimal sensor combination to obtain a feature data matrix; Performing a deep Q network analysis on the characteristic data matrix and the sensor response change rate to generate a fault diagnosis trigger signal; According to the fault diagnosis trigger signal, the characteristic data matrix is ​​subjected to sensor channel block mapping and position encoding to obtain a time-space separation characteristic vector; The time-space separation feature vector is input into a bidirectional long short-term memory network to determine the fault type, and the fault type determination result is output.

[0006] A second aspect of the present application provides an intelligent fault diagnosis system for a multi-channel gas sensor, the intelligent fault diagnosis system for the multi-channel gas sensor comprising: A construction module is used to perform random geometric analysis and response restriction area construction on the response data of n sensor channels in a multi-channel gas sensor array to obtain an effective sensor set; A screening module, used for performing correlation analysis and sensitivity precision dilution measurement function screening on the response data between the sensor channels in the effective sensor set to obtain the optimal sensor combination; An extraction module, used for periodically sampling and extracting features from the response data of the optimal sensor combination to obtain a feature data matrix; An analysis module, used for performing a deep Q network analysis on the characteristic data matrix and the sensor response change rate to generate a fault diagnosis trigger signal; An encoding module, used for performing sensor channel block mapping and position encoding on the characteristic data matrix according to the fault diagnosis trigger signal to obtain a time-space separation characteristic vector; The determination module is used to input the time-space separation feature vector into a bidirectional long short-term memory network to determine the fault type and output a fault type determination result.

[0007] The third aspect of the present application provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the electronic device executes the above-mentioned intelligent fault diagnosis method for the multi-channel gas sensor.

[0008] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned intelligent fault diagnosis method for a multi-channel gas sensor.

[0009] Compared with the prior art, the present application has the following beneficial effects: through random geometric analysis and response restriction area construction, a set of effective sensors with stable states is dynamically screened out, and at the same time, faulty sensors are timely identified and isolated, which significantly improves the reliability of the system; the sensitivity precision dilution measurement function is used for sensor optimization, which reasonably reduces the number of redundant sensors, reduces the computational complexity, and ensures the accuracy of fault diagnosis; the fault triggering mechanism based on the deep Q network and the diagnostic utility function realizes the intelligent triggering of fault diagnosis, avoiding the waste of computing resources caused by the traditional fixed-cycle diagnosis; the temporal attention mechanism and position encoding strategy are introduced to effectively extract the spatiotemporal characteristics of the sensor, and make full use of the spatial distribution information and temporal variation law of the sensor; the fault diagnosis method using a bidirectional long short-term memory network combined with a feature memory module realizes the accurate identification and classification of different types of faults, and can give the specific type, location and severity of the fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0011] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0012] Figure 1 It is a flowchart of an intelligent fault diagnosis method for a multi-channel gas sensor provided by an embodiment of the present invention; Figure 2 is a schematic block diagram of the structure of an intelligent fault diagnosis system for a multi-channel gas sensor provided by an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0015] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0016] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 , an embodiment of the intelligent fault diagnosis method of the multi-channel gas sensor in the embodiment of the present application includes: Step 100, performing random geometric analysis and response restriction region construction on response data of n sensor channels in a multi-channel gas sensor array to obtain an effective sensor set; It is understandable that the execution subject of the present application may be an intelligent fault diagnosis system for a multi-channel gas sensor, or a terminal or a server, which is not limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0017] Specifically, the response data of n sensor channels in the multi-channel gas sensor array are collected. Where n represents the total number of sensor channels, and the collected data is the response voltage value of each sensor channel in a specific time period to form a response value sequence (i=1,2,...,n), where each Represents the response voltage of the i-th sensor. The data is input to the data processing unit for calculation. The response value sequence of all sensor channels is accumulated and summed, and then the result is divided by the number of sensor channels n to obtain the response mean. . Response mean The average voltage value representing the response of all sensors is a benchmark value reflecting the performance of the entire sensor array. , and compare it with the mean Perform difference calculation to obtain difference sequence , and perform square sum operation on the difference sequence, and finally divide by the square root , and the response standard deviation is , which measures the discreteness of all sensor response data. Based on the response mean and the response standard deviation , perform boundary value calculations and set the upper and lower limits of the response limit area. Set to , lower limit Set to , which represents the normal working response range. Through this limited area, it is judged whether the response value of each sensor is within a reasonable range. Compare the size of each response value in satisfy , then the response value is considered to be within the normal working range and marked as a valid response value. At this time, the response value marking result is represented by binary, 1 means valid, and 0 means invalid. According to the response value marking result, the n sensor channels are classified. The sensor channels marked as 1 are valid sensors, and their responses are within the predetermined range, indicating that these sensors are working normally; while the sensor channels marked as 0 are classified into the set of sensors to be diagnosed. The response values ​​of these sensors to be diagnosed fail to meet the normal working range and require further analysis and diagnosis. For the sensors in the valid sensor set, the response stability is calculated. The response changes of each sensor in multiple consecutive sampling cycles are evaluated to determine whether it is stable and reliable. The coefficient of variation CV of the response value of each sensor in 10 consecutive sampling cycles is calculated, and the coefficient of variation is equal to the standard deviation of the response value in the time window divided by the mean of the response value. The smaller the coefficient of variation CV, the more stable the response of the sensor, otherwise it means that its response has large fluctuations. The obtained coefficient of variation CV is compared with the preset first threshold Th, and the threshold Th is set to 0.1. If the coefficient of variation of a sensor is lower than the threshold, the sensor response is considered stable and it will continue to be retained in the effective sensor set; if the coefficient of variation is higher than the threshold, the sensor response is considered unstable and needs to be transferred to the sensor set to be diagnosed. For the sensor channels in the sensor set to be diagnosed, the fault type is determined to obtain the specific fault type of each sensor to be diagnosed, thereby achieving accurate fault diagnosis.

[0018] Step 200: performing correlation analysis and sensitivity precision dilution measurement function screening on the response data between the sensor channels in the effective sensor set to obtain the optimal sensor combination; Specifically, the response data between two sensor channels in the effective sensor set are subjected to correlation analysis. The Pearson correlation coefficient is calculated to measure the linear relationship between the response data of the two sensor channels. The value range of the Pearson correlation coefficient is from -1 to 1, where the closer the value is to 1, the more similar the responses of the two sensors are, the closer the value is to -1, the opposite direction of the responses of the two sensors are, and the closer the value is to 0, the more there is no significant linear relationship between them. By calculating the Pearson correlation coefficient between each pair of sensor channels in the effective sensor set, a preliminary correlation coefficient matrix is ​​obtained, which reflects the degree of correlation between the responses of all sensors. The initial correlation coefficient matrix is ​​normalized to map each correlation coefficient value between 0 and 1. A weighted sum operation is performed on each row element in the standardized correlation coefficient matrix, and the result is divided by the number of sensor channels n to obtain a weight coefficient sequence, which reflects the relative importance of each sensor channel in the entire sensor array. The weight coefficient sequence is arranged in a diagonal matrix form to form a sensor weight matrix. The weight matrix represents the relative weights between different sensors and the correlation between them. At the same time, the covariance of the response data in the effective sensor set is calculated to obtain the covariance matrix of the response data. The covariance matrix is ​​used to measure the degree of common variation of response data between sensor channels and reflect the consistency of response fluctuations between different sensors. The covariance matrix reflects the interaction between sensors and their common influence on fault detection results. Based on the sensor weight matrix and covariance matrix, the sensitivity dilution of precision metric function (SDoP value) is calculated. The SDoP value is an indicator to measure the effectiveness of the sensor combination. Its calculation formula is the trace of the product of the weight matrix and the inverse matrix of the covariance matrix. The smaller the SDoP value, the higher the sensitivity and accuracy of the sensor combination, and the better the utility of the combination. By calculating the SDoP value, an evaluation index for the sensor combination is obtained, which helps to screen out sensor combinations with better performance. The SDoP value is compared with the second preset threshold. If the SDoP value of a certain sensor combination is less than the second preset threshold, it means that the combination performs well in terms of sensitivity and accuracy, and it is marked as a candidate sensor combination. The redundancy index is calculated for each combination in the candidate sensor combination set. Redundancy reflects the repeatability between multiple sensor channels. Combinations with lower redundancy mean that they will not generate redundant information when detecting faults, thereby improving the efficiency of the system. By calculating the redundancy of each candidate combination, the combination with the smallest redundancy and the number of sensors not less than the minimum value set by the system is selected as the optimal sensor combination.

[0019] Step 300: Periodically sample and extract features of the response data of the optimal sensor combination to obtain a feature data matrix; It should be noted that the data of each sensor channel in the optimal sensor combination is sampled periodically. Data sampling is performed on each sensor channel for 10 seconds, and a set of response signals is obtained each time sampling. Over time, an original sampling data set is accumulated. The response signal represented by each sampling point is the voltage or other related physical quantity of the sensor at that moment, thereby reflecting the response change of the sensor under the working state. The original sampling data set is standardized. Based on the maximum value in the data of each sensor channel and minimum value The response data is standardized, and the range of all data is normalized to between 0 and 1 to obtain a normalized response data set. The standardized data set is divided into multiple data windows. The length of the data window is set to 100 data sampling points, that is, each data window contains 100 time steps of response signal data. In order to enable the data window to cover the entire sampling data set, the sliding interval between adjacent windows is set to 10 data sampling points. Each time the window slides, it only advances 10 data points to generate multiple adjacent and overlapping data windows. Through this sliding window method, smaller time changes and trends are captured in the time series. For each segmented data window, the time domain statistics are calculated for the data therein, including mean, variance, peak and valley values. These time domain features can describe the basic statistical characteristics of the sensor response signal in a specific time window. For example, the mean can reflect the central position of the response, the variance describes the discreteness of the data, and the peak and valley values ​​reveal the extreme values ​​in the signal, helping to identify abnormal fluctuations. The response characteristics in each data window are obtained through the time domain feature parameters. Based on the time domain feature parameter set, the distribution characteristics of the data in each window are calculated, including the skewness and kurtosis values ​​of the data. The skewness value is used to measure the symmetry of the data distribution, while the kurtosis value measures the sharpness of the data distribution. By calculating these distribution characteristics, more details of the data in the time domain are captured and potential signal anomalies or fault modes are identified. The calculation of skewness and kurtosis can reveal the asymmetry or extreme value concentration of the signal, which is helpful for fault diagnosis and pattern recognition. The time domain feature parameter set and the distribution feature parameter set are merged to obtain a comprehensive feature vector of a single time window, which contains the statistical characteristics and distribution characteristics of the signal in the window. All feature vectors are combined into a feature vector group. Time series features are constructed based on the feature vector group. The feature vectors of M consecutive time windows are arranged and combined in time order, where M represents the number of time windows. By arranging the feature vectors in time order, a time series feature matrix is ​​constructed to reflect the dynamic characteristics of the signal changing over time. The time series feature matrix is ​​reorganized to integrate the feature data of the same sensor channel. The feature data of each sensor in different time windows are summarized so that the time series response characteristics of each sensor are systematically displayed. These integrated feature data constitute a feature data matrix.

[0020] Step 400, performing a deep Q network analysis on the characteristic data matrix and the sensor response change rate to generate a fault diagnosis trigger signal; Specifically, the response data in the feature data matrix is ​​differentially calculated between adjacent time windows to capture the changing trend of the sensor response data over time. By calculating the difference between adjacent time windows, the amplitude of the sensor response change in each time window is obtained. These differences reflect the dynamic changes of the sensor and help reveal the abnormal or fault mode of the sensor. The results of the differential calculation are normalized according to the sampling time interval to ensure that the differences between different time windows have the same dimension and can be compared within a unified range to obtain the sensor response change rate sequence. The feature data matrix and the sensor response change rate sequence are data fused to construct the system state vector. The system state vector contains all the statistical characteristic parameters of the sensor and their change rate information. By fusing static features and change rate features, the current state of the system is described and sufficient basis is provided for the subsequent decision-making process. The action space is set for the input state matrix. The design of the action space is based on the requirements of the fault diagnosis system. Each action represents a possible operation of the system in a given state. For example, an action corresponds to marking a sensor as a fault state, and another action corresponds to further checking a sensor. After setting the action space, each action is assigned a unique identification code and encoded in the deep Q network. A deep Q network is constructed based on the input state matrix and action space encoding. In the deep Q network, the convolution layer and the pooling layer are used to extract and reduce the input state features so as to extract the most valuable features from the complex input data. The convolution layer can extract useful spatial features from the local area, while the pooling layer reduces the amount of calculation by reducing the dimension and retains the main information. After being processed by these layers, the state feature encoding is obtained, which contains the main features of the sensor data and can be passed in the network for decision-making. The Q value of each action is calculated using a fully connected layer. The Q value represents the long-term reward that can be obtained by taking a specific action in the current state. Through the greedy strategy, the network selects the action with the largest Q value, that is, the action that is most likely to lead to normal operation of the system or successful fault diagnosis in the current state. Based on the selection process, the state-action decision value is obtained, which represents the best decision of the deep Q network for the current state. In order to evaluate the effect of the diagnostic decision, a diagnostic utility function is constructed. This function weightedly combines the inverse of the end-to-end diagnostic delay time and the diagnostic accuracy to form a comprehensive evaluation index. The inverse of the diagnostic delay time reflects the timeliness of the diagnostic process, while the diagnostic accuracy reflects the reliability of the fault diagnosis. By combining these two indicators, the comprehensive utility of each diagnostic decision is evaluated to determine whether it meets the performance requirements of the system. The evaluation value of the utility function is compared with the third preset threshold. When the utility function evaluation value is lower than the threshold, the system believes that the current diagnostic decision is not effective and a new diagnostic process needs to be triggered. At this time, a diagnostic trigger flag is generated, and a fault diagnosis trigger signal is generated according to the trigger flag.The signal contains information such as trigger time, state number and trigger type, which serves as diagnostic feedback for the system and is used for subsequent fault handling or alarm.

[0021] Step 500: According to the fault diagnosis trigger signal, the characteristic data matrix is ​​subjected to sensor channel block mapping and position encoding to obtain a time-space separation characteristic vector; Specifically, feature block segmentation is performed on each sensor channel in the feature data matrix. The response data of each sensor channel is divided into multiple feature blocks, where each feature block contains the time domain features and distribution features of the sensor channel in a certain time period. The time domain feature parameter set includes statistics such as mean, variance, peak, and valley, while the distribution feature parameter set includes distribution indicators such as skewness and kurtosis. By grouping these parameters, a feature block sequence of each sensor channel is formed. At the same time, a unique physical space identifier is assigned to each sensor channel based on the feature block sequence to record the relative position of the sensor in the array. The identifier gives each sensor channel a spatial position attribute. The feature block sequence of each sensor channel is matrix-joined with its position encoding data to form a combined feature matrix containing both sensor response feature information and position information. The initial spatiotemporal feature matrix contains the dynamic response data of the sensor and also combines the physical position of the sensor in the array. The data in the initial spatiotemporal feature matrix is ​​segmented. The length of each time segment is set to 100 sampling points, and each time segment contains sensor response data within a certain time range. By segmentation, the entire spatiotemporal feature matrix is ​​converted into multiple temporal block matrices, each of which represents the sensor response feature within a time window. The temporal block matrix is ​​input into the temporal attention network to calculate the feature correlation score in the temporal dimension. The temporal attention network highlights the temporal features that are most critical to fault diagnosis by weighting the features of different time periods. By calculating the attention weight coefficient of each time segment, the network can automatically identify which temporal features have a greater influence in the diagnosis process, thereby giving these features a higher weight. Based on the temporal attention weight coefficient, the temporal block matrix is ​​weighted, so that the network can pay more attention to the key temporal features in the diagnosis process, improving the accuracy and efficiency of diagnosis. The weighted temporal feature matrix is ​​dimensionally integrated, the two-dimensional feature matrix is ​​compressed into a one-dimensional feature sequence, and the temporal information is condensed into a more concise feature vector. The compressed feature vector is normalized, all eigenvalues ​​are uniformly normalized to the interval [0,1], the scale differences between different features are eliminated, and all features are compared and analyzed under the same standard. After the above steps, the spatiotemporal separation feature vector is obtained.

[0022] Step 600: Input the time-space separation feature vector into a bidirectional long short-term memory network to determine the fault type, and output the fault type determination result.

[0023] Specifically, the spatiotemporal separation feature vectors are input into the bidirectional long short-term memory network forward and backward according to the time series. Through the bidirectional long short-term memory network, the forward and backward information of the sequence data are used for feature learning at the same time. The forward propagation can process the historical data of the current moment and before, while the backward propagation can capture the impact of the future moment, thereby fully considering the bidirectional dependency of the time series. Through this step, the network can more comprehensively understand the temporal characteristics of the data and generate a bidirectional feature representation sequence. Based on the bidirectional feature representation sequence, a feature memory queue is constructed. The length of the queue is set to 1000 feature patterns. In the update process of the queue, the first-in-first-out principle is adopted, that is, whenever a new feature pattern enters, the oldest feature pattern will be discarded. The queue update mechanism ensures that the system can always pay attention to the latest sensor data, so as to maintain an accurate judgment of the current state. With the accumulation of data, the memory queue forms a dynamic feature memory library, which stores the historical feature patterns of the sensor. In order to improve the diagnostic efficiency, the feature patterns stored in the dynamic feature memory library are clustered. Through cluster analysis, feature patterns with high similarity are identified and classified into the same category to obtain a normal working feature pattern set, which represents the typical response characteristics of the sensor under normal working conditions. The bidirectional feature representation sequence is feature matched with the normal working feature pattern set. The Euclidean distance between the bidirectional feature representation sequence and the normal working feature pattern set is calculated as the distance deviation between the features. The Euclidean distance can measure the similarity between two feature vectors. The larger the deviation, the more obvious the difference between the current sensor response and the normal working mode. These calculation results constitute a feature deviation matrix, and each element in the matrix represents the difference between the feature and the normal mode at a certain moment. According to the deviation value, a fault type discrimination threshold is set, and the features with a deviation value greater than the fifth preset threshold are marked as potential fault features. Through this marking, features that indicate faults are identified from a large amount of sensor data. Based on the fault feature marking results, fault mode classification is performed. Fault modes with similar deviation characteristics are classified into the same category, so that the system can classify different fault phenomena into similar patterns and identify potential fault types. Through this classification method, the specific fault type of the equipment is gradually determined, and each type of fault is analyzed. Fault level classification is performed for each initial fault type. The fault level is divided into three levels: slight, moderate and severe based on the size of the characteristic deviation value. Combined with the initial fault type and fault level classification results, a complete fault type determination result is generated. The result includes information such as fault type, fault level and fault location.

[0024] In the embodiments of the present application, through random geometric analysis and response restriction area construction, a set of effective sensors with stable states is dynamically screened out, and faulty sensors are timely identified and isolated, which significantly improves the reliability of the system; a sensitivity precision dilution metric function is used for sensor optimization, which reasonably reduces the number of redundant sensors, reduces the computational complexity, and ensures the accuracy of fault diagnosis; a fault trigger mechanism based on a deep Q network and a diagnostic utility function realizes intelligent triggering of fault diagnosis, avoiding the waste of computing resources caused by traditional fixed-cycle diagnosis; a temporal attention mechanism and a position encoding strategy are introduced to effectively extract the spatiotemporal characteristics of the sensor, and make full use of the spatial distribution information and temporal variation law of the sensor; a fault diagnosis method using a bidirectional long short-term memory network combined with a feature memory module is used to realize accurate identification and classification of different types of faults, and can give the specific type, location and severity of the fault.

[0025] In a specific embodiment, the process of executing step 100 may specifically include the following steps: Response data of n sensor channels of the multi-channel gas sensor array are collected, where n represents the total number of sensor channels, and a response value sequence Ri (i=1, 2, ..., n) is obtained, where Ri represents the response voltage value of the i-th sensor channel; The response value sequence Ri is input into the data processing unit for cumulative summation operation, and the sum is divided by the number of sensor channels n to obtain the response mean μ; The difference between each response value in the response value sequence Ri and the response mean μ is calculated and recorded as the difference sequence Di. The difference sequence Di is squared and then squared and divided by n-1 to obtain the response standard deviation σ. The boundary value is calculated based on the response mean μ and the response standard deviation σ, the upper limit UL is set to μ+2σ, and the lower limit LL is set to μ-2σ, and the boundary value of the response limit area is obtained; Compare each response value in the response value sequence Ri with the boundary value of the response limit area. When the response value Ri satisfies LL≤Ri≤UL, it is marked as a valid response value to obtain a response value marking result. The response value marking result is represented by a binary number, 1 represents valid, and 0 represents invalid. Classify the n sensor channels according to the response value marking results, classify the sensor channels marked as 1 into the valid sensor set VS, and classify the sensor channels marked as 0 into the sensor set to be diagnosed DS; The response stability of the sensor channels in the effective sensor set VS is calculated, and the coefficient of variation CV of the response value within 10 consecutive sampling periods is calculated, where the coefficient of variation CV is equal to the standard deviation within the time window divided by the mean, and the stability evaluation value is obtained. The stability evaluation value is a dimensionless number; The stability evaluation value is compared with a first preset threshold value Th, wherein the first preset threshold value Th is set to 0.1, and the sensor channels whose coefficient of variation CV is lower than the first preset threshold value Th are retained in the effective sensor set VS, and the sensor channels whose coefficient of variation CV is higher than Th are transferred to the sensor set to be diagnosed DS, and the fault type of the sensor channels in the sensor set to be diagnosed is directly determined to obtain the fault determination result of the sensor set to be diagnosed.

[0026] Specifically, the response data is obtained from each sensor channel. The number of channels in the sensor array is set to The response data of each channel is collected through the sensor system. The response data of each sensor channel is represented as a series of voltage values, recorded as ,in indicates the number of the sensor channel, and It is The response voltage values ​​of the sensor channels. These data form a response value sequence: ; The response values ​​are input into the data processing unit for statistical analysis. The response value sequence is accumulated and summed, and the sum of the response voltage values ​​of all sensor channels is calculated, and then the sum is divided by the number of sensor channels. , get the mean of the response values The formula is as follows: ; in, Represents the mean of the response value sequence, reflecting the overall response level of the multi-channel gas sensor array. The calculation of the mean helps to measure the average level of the response of all sensor channels. Calculate the standard deviation of the response for each sensor channel , and get the fluctuation of the response value. Calculate the difference between each response value and the mean, and record the difference sequence as ,Right now: ; For difference sequence Square it, get the sum of the squares, take the square root, and divide by To get the standard deviation , the formula is: ; Reflects the discreteness of the response value, that is, the response volatility of the sensor channel. The larger the standard deviation, the greater the volatility of the response. and the response standard deviation , calculate the boundary value to determine which sensor response values ​​are valid. Based on the standard deviation, add or subtract twice the standard deviation to get the upper and lower limits of the response value. and lower limit They are calculated as: ; ; The boundary value is used to limit the response range of the sensor. falls within this range, that is, , then the response value is considered valid, otherwise it is invalid. Mark the response value of each sensor according to whether the response value is within the valid range. satisfy , it is marked as a valid response value, that is, expressed in binary, 1 means valid, 0 means invalid. Response value marking result It is expressed as: ; All sensor channels are classified according to the response value marking results. Sensor channels marked as 1 belong to the valid sensor set. The sensor channel marked as 0 enters the set of sensors to be diagnosed. After classification, the effective sensor set Calculate the response stability of the sensor channels in the . Calculate the coefficient of variation of the response value of each sensor channel within 10 consecutive sampling cycles , coefficient of variation It is calculated by dividing the standard deviation of the response by the mean, using the following formula: ; in, is the standard deviation of the response values ​​within 10 cycles, is the response mean value within these 10 cycles. By calculating the coefficient of variation, the response stability of the sensor channel is evaluated. The smaller the coefficient of variation, the more stable the sensor response. With the preset first threshold h is compared, where h is set to 0.1. If the coefficient of variation of a sensor channel is lower than the threshold, it indicates that the response of the sensor channel is stable and the sensor is retained in the valid sensor set. Medium; if the coefficient of variation is above the threshold h, the response of the sensor channel is unstable, and it is transferred to the sensor set to be diagnosed. For the sensor channels in the set of sensors to be diagnosed, direct fault determination is performed according to their characteristics, and the fault determination result is output.

[0027] In a specific embodiment, the process of executing step 200 may specifically include the following steps: The Pearson correlation coefficient is calculated for the response data between two sensor channels in the effective sensor set, where the correlation coefficient value ranges from -1 to 1, and an initial correlation coefficient matrix is ​​obtained; The initial correlation coefficient matrix is ​​normalized, and the correlation coefficient value is mapped to between 0 and 1 to obtain a standardized correlation coefficient matrix. The weighted sum operation is performed on each row element in the standardized correlation coefficient matrix, and the weight coefficient sequence is obtained by dividing it by the number of sensor channels. Arrange the weight coefficient sequence into a diagonal matrix form to obtain the sensor weight matrix, and perform covariance calculation on the response data in the effective sensor set to construct the covariance matrix of the response data; Based on the sensor weight matrix and covariance matrix, the sensitivity dilution of precision metric function SDoP value is calculated, where the SDoP value is the trace of the product of the weight matrix and the inverse matrix of the covariance matrix, and the sensor combination evaluation index is obtained; The sensor combination evaluation index is compared with a second preset threshold value, and when the SDoP value is less than the second preset threshold value, the corresponding sensor combination is marked as a candidate combination to obtain a candidate sensor combination set; The redundancy index is calculated for each combination in the candidate sensor combination set, and the combination with the smallest redundancy and the number of sensors not less than the minimum value set by the system is selected as the optimal sensor combination.

[0028] Specifically, the Pearson correlation coefficient is calculated for the response data between two sensor channels in the effective sensor set. The Pearson correlation coefficient is a statistic that measures the linear correlation between two variables (response data between sensor channels). and , and its Pearson correlation coefficient Calculated by the following formula: ; in, and Respectively represent and The sensor in The response data of the sampling points, and Respectively represent and The average value of sensor channels, is the total number of sampling points of the response data. The value range is from -1 to 1, where a value of 1 indicates a perfect positive correlation, a value of -1 indicates a perfect negative correlation, and a value of 0 indicates no linear relationship. Select sensor channels, calculate the Pearson correlation coefficient between each pair of sensor channels, and form an initial correlation coefficient matrix. Assume that the effective sensor set contains sensor channels, the initial correlation coefficient matrix is a A matrix where each element Indicates The sensor channel and The Pearson correlation coefficient between the sensor channels. Normalize the initial correlation coefficient matrix. Map the correlation coefficient value to the [0,1] interval so that subsequent operations are not affected by negative values, which facilitates the calculation of weight coefficients. The normalization formula is as follows: ; in, is the normalized correlation coefficient, is the original Pearson correlation coefficient. After normalization, all correlation coefficient values will be in the range of 0 to 1. After that, the weighted sum of each row is calculated to obtain the weight coefficient sequence. Assume that the standardized correlation coefficient matrix is , the weighted sum of each row is calculated by the following formula: ; in, Indicates The weight coefficient of each sensor channel, is the element in the normalized correlation coefficient matrix. The result of the weighted sum operation is divided by the number of sensor channels , and obtain a weight coefficient sequence , where each Representative The weight coefficient sequence is arranged into a diagonal matrix ; The diagonal matrix Reflects the relative importance of each sensor channel. Using the weight matrix , combined with an effective sensor set Calculate the covariance matrix of the response data. The formula for calculating the covariance matrix is ​​as follows: ; in, Each column of the response data matrix represents The response data of the sampling points, is the mean vector of the response data. The covariance matrix Describes the linear relationship and volatility of response data between different sensor channels. Based on the sensor weight matrix and the covariance matrix , calculate the value of the sensitivity dilution of precision metric function SDoP. This metric function is used to evaluate the sensitivity and precision of the sensor combination. The calculation formula of SDoP is: ; in, represents the trace operation of a matrix, is the inverse matrix of the covariance matrix. The calculated SDoP value represents the performance of the sensor combination. A lower SDoP value means that the sensor combination has better diagnostic performance. The calculated SDoP value is compared with the second preset threshold When SDoP is less than , the sensor combination is considered as a candidate combination and marked as a valid candidate combination, and the candidate sensor combination set is obtained. . The redundancy index is calculated for each combination in the candidate sensor combination set. The redundancy index reflects the redundant information between sensor channels. A combination with lower redundancy means that the information between its sensor channels is more complementary. The redundancy index is calculated by the following formula: ; After calculating the redundancy, select the combination with the smallest redundancy and the number of sensors not less than the minimum value set by the system as the optimal sensor combination. Assume that the minimum number of sensors set by the system is , then the final optimal sensor combination is the optimal combination selected from the candidate combinations, with the minimum redundancy and the number of sensors being greater than or equal to .

[0029] In a specific embodiment, the process of executing step 300 may specifically include the following steps: The sensor channels in the optimal sensor combination are sampled for a period of 10 seconds, and a set of response signals is obtained each time the sampling is performed to obtain the original sampling data set; Based on the maximum value Vmax and the minimum value Vmin in the original sampling data set, the response data is normalized, and the data range is normalized to the interval of 0 to 1 to obtain a normalized response data set; A data window is set for the normalized response data set, the window length is set to 100 data sampling points, and the sliding interval between adjacent windows is set to 10 data sampling points to obtain a segmented data window set; Calculate the time domain statistics of the data in each window of the segmented data window set, including mean, variance, peak value, and valley value, and obtain a time domain feature parameter set; Based on the time domain feature parameter set, the distribution characteristics of the data in each window are calculated, including the skewness value and the kurtosis value of the data distribution, to obtain the distribution feature parameter set, and the time domain feature parameter set and the distribution feature parameter set are feature merged to form a comprehensive feature vector of a single time window, to obtain a feature vector group; The time series features are constructed based on the feature vector group. The feature vectors of M consecutive time windows are arranged and combined in chronological order, where M is the number of time windows, to obtain a time series feature matrix. The time series feature matrix is ​​reorganized according to the sensor channel identifier, and the feature data of the same sensor channel are integrated together to obtain a feature data matrix.

[0030] Specifically, each sensor channel in the optimal sensor combination is sampled periodically, once every 10 seconds, and each sampling obtains a data point as a response signal to obtain an original sampling data set. Assume that there is an optimal sensor combination, including sensor channels, with a sampling period of 10 seconds. The response data of each sensor is expressed as: ; in, It is The response data sequence of each sensor, is the total number of sampling points of the data. The sampled data set contains all the response signals of each sensor channel. Based on the maximum value in the original sampled data set and minimum value , standardize the response data. The response data of each sensor channel is normalized to the interval [0,1] to eliminate the dimension differences of different sensors and ensure the consistency of the data. The standardization formula is as follows: ; in, It is The sensor in The response value after standardized processing at each time point is is the original response data, and are the maximum and minimum values ​​in the response data of the channel, respectively. Set the data window for the normalized response data set to facilitate the subsequent time domain feature extraction. Assume that the response data sequence is divided into multiple data windows with a length of 100 sampling points, and the sliding interval of each window is 10 sampling points. The windows are generated as follows: the first window contains the 1st to 100th data points, the second window contains the 10th to 109th data points, and so on. The data set of each window is represented as follows: ; in, Indicates The response data of each window contains 100 consecutive sampling points. The original response data is divided into multiple windows by the sliding window method to generate a set of segmented data windows. For each data window, the time domain statistics are calculated, mainly including statistical indicators such as mean, variance, peak and valley. These characteristics can effectively reflect the fluctuation and trend of the signal. Based on the time domain characteristic parameters, the distribution characteristics of each window data are calculated, including skewness and kurtosis. Skewness measures the symmetry of data distribution, and the calculation formula is: ; in, and are the mean and standard deviation of the window respectively. The skewness value reflects the degree of skewness of the data distribution. A positive value indicates right skewness, a negative value indicates left skewness, and 0 indicates symmetrical distribution. Kurtosis measures the sharpness of the data distribution. The calculation formula is: ; in, is the standard deviation of the window data. The higher the kurtosis value, the sharper the data distribution, and the lower the value, the flatter the data distribution. By calculating distribution features such as skewness and kurtosis, we can enhance our understanding of data distribution features. The combination of time domain features and distribution features constitutes a complete feature parameter set. The time domain feature parameter set and the distribution feature parameter set are combined to obtain the comprehensive feature vector of each time window. windows, comprehensive feature vector It is expressed as: ; These feature vectors can comprehensively characterize the response data characteristics within the time window at a lower dimension. For multiple time windows, arrange them in chronological order to obtain the time series feature matrix , which is expressed as: ; in, is the number of time windows. Time series feature matrix Contains the time window from the first to the All feature data of a time window. The time series feature matrix is ​​reorganized according to the sensor channel, and the feature data of the same sensor channel are integrated together to obtain the final feature data matrix Assume that sensor channels, reorganize the feature matrix of each sensor channel according to the channel identifier, and finally obtain the feature data matrix Contains characteristic data for all sensors.

[0031] In a specific embodiment, the process of executing step 400 may specifically include the following steps: The response data in the characteristic data matrix is ​​differentially calculated in adjacent time windows and normalized according to the sampling time interval to obtain the sensor response change rate sequence; The characteristic data matrix and the sensor response change rate are fused to construct the system state vector, which contains all statistical characteristic parameters and their change rate information, and obtain the input state matrix of the deep Q network; Set the action space for the input state matrix and assign a unique identification code to each action to obtain the action space encoding; A deep Q network is constructed based on the input state matrix and action space encoding, and the state features are extracted and reduced in dimension through convolutional layers and pooling layers to obtain the state feature encoding; The state feature encoding is used to calculate the Q value of each action through the fully connected layer, and the action with the largest Q value is selected based on the greedy strategy to obtain the state-action decision value; A diagnostic utility function is constructed for the state-action decision value, and the inverse of the end-to-end diagnostic delay time and the diagnostic accuracy are weighted to obtain the utility function evaluation value; The utility function evaluation value is compared with a third preset threshold value. When the utility function evaluation value is lower than the third preset threshold value, a new diagnostic process is triggered, a diagnostic trigger flag is obtained, and a fault diagnosis trigger signal including a trigger time, a state number and a trigger type is generated according to the diagnostic trigger flag.

[0032] Specifically, the response data is obtained from the feature data matrix, and the difference calculation of adjacent time windows is performed on these data. The difference calculation can capture the dynamic characteristics of data changes and identify the response change trend of the sensor. Assume that the feature data matrix Include The data of each time window is ,in is the number of features in each time window, Indicates the number of the time window. The response change rate sequence is obtained by calculating the difference between adjacent time windows. : ; The data after difference Indicates the amount of change between adjacent time windows, which can help the system analyze the dynamic changes of data and reveal the change pattern of sensor status. In order to eliminate the influence of time interval, the differential data is normalized. Assume that the sampling time interval is , the rate of change is normalized using the following formula: ; The processed response change rate sequence is Normalized It can reflect the change speed of each sensor channel in unit time, eliminate the influence of time interval, and make the change data of different time windows uniformly compared. The characteristic data matrix and the sensor response change rate are fused to construct the system state vector. The system state vector is the input of the deep Q network, which contains all statistical characteristic parameters and their change rate information. Assume that The data of time windows, each window contains time domain features and the corresponding rate of change , the system state vector of each time window is expressed as: ; Combine the system state vectors of all time windows to get the input state matrix of the deep Q network , expressed as: ; This state matrix contains the statistical characteristics and change rate of each time window, which can fully reflect the current state and change trend of the sensor, and serve as the input data for subsequent deep Q network training. After constructing the input state matrix, define the action space of the deep Q network and assign a unique identification code to each action. The action space represents the different operations that the system can choose, such as adjusting sensor thresholds, resampling, performing fault diagnosis, etc. Assume that the size of the action space is , each action Corresponding to a unique identification code , expressed as: ; For each action, a unique code is given, and a deep Q network is used to learn how to select the optimal action under different system states. A deep Q network is constructed based on the input state matrix and action space encoding. In the construction of the network, convolutional layers and pooling layers are used to extract and reduce the dimension of state features. The convolutional layer can extract local features in the input state matrix, and the pooling layer can reduce the dimension and reduce the amount of calculation, thereby improving the training efficiency and prediction performance of the model. In this way, the deep Q network can effectively learn the complex relationship between the system state and action, and extract the key features that affect the decision. After processing by the convolutional layer and the pooling layer, the Q value of each action is calculated through the fully connected layer. Q value is the core concept in the deep Q network, which represents the expected reward that can be obtained by performing an action in a specific state. The calculation formula of Q value is: ; Indicates the current status. Indicates the currently selected action. is in state Next select action Immediate returns, is a discount factor, indicating the importance of future rewards, is the maximum Q value of all possible actions in the next state. By training the deep Q network, the model can learn the optimal state-action mapping and thus select the most appropriate action. Based on the Q value calculated above, the greedy strategy is used to select the action with the largest Q value as the current decision. The greedy strategy can select an action that maximizes the expected return, that is, the action with the largest Q value. The formula for the greedy strategy selection is: ; in, represents the optimal action, and the action with the maximum Q value is selected to help the system make the best decision. After obtaining the state-action decision value, a diagnostic utility function is constructed, which weights the inverse of the end-to-end diagnostic delay time and the diagnostic accuracy to evaluate the utility of the current decision. The diagnostic delay time reflects the time required for the system to make a fault diagnosis, and the diagnostic accuracy indicates the correctness of the fault diagnosis. The formula of the utility function is: ; in, and is a weighting coefficient that controls the relative importance of delay time and accuracy. The utility function can quantify the comprehensive effect of each action and provide a basis for decision-making. The utility function evaluation value is compared with the third preset threshold to trigger a new diagnostic process. When the evaluation value of the utility function is lower than the third preset threshold, a new diagnostic process is triggered, and a fault diagnosis trigger signal containing the trigger time, state number and trigger type is generated. This signal is used for subsequent system adjustment or alarm processing. The specific judgment formula is: .

[0033] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Each sensor channel in the feature data matrix is ​​segmented into feature blocks, grouped according to the time domain feature parameter set and the distribution feature parameter set to obtain a feature block sequence, and a unique physical space identifier is assigned to each sensor channel based on the feature block sequence, and the relative position information in the sensor array is recorded to obtain position coding data; The feature block sequence and the position coding data are concatenated by matrix operation to construct a combined feature matrix containing both feature information and position information to obtain an initial spatiotemporal feature matrix. The data in the initial spatiotemporal feature matrix is ​​then divided into time segments of fixed size, each of which is 100 sampling points long, to obtain a time series block matrix. The temporal block matrix is ​​input into the temporal attention network, the feature correlation score on the temporal dimension is calculated, the temporal attention weight coefficient is obtained, and the temporal block matrix is ​​weighted according to the temporal attention weight coefficient to highlight the key temporal features and obtain the weighted temporal feature matrix; All time series features in the weighted time series feature matrix are dimensionally integrated, the two-dimensional feature matrix is ​​compressed into a one-dimensional feature sequence to obtain a compressed feature vector, and the compressed feature vector is normalized to normalize all eigenvalues ​​to the interval [0,1] to obtain a time-space separation feature vector.

[0034] Specifically, feature blocks are segmented for each sensor channel in the feature data matrix. The feature data matrix contains the response data collected by each sensor over a period of time. Assuming that there are N sensor channels in the matrix, the data set of each channel is processed in blocks according to time domain features and distribution features. Among them, the time domain feature parameter set includes information such as response mean, variance, peak value, valley value, etc., and the distribution feature parameter set includes the skewness and kurtosis of the data, reflecting the distribution characteristics of the data. For the data of each sensor channel, it is grouped according to the time domain features and distribution features to generate a feature block sequence. Each feature block is regarded as a small set containing multiple feature parameters. Assume that the feature block sequence of the i-th sensor channel is: ; in, is the set of time-domain features of the sensor, such as mean and variance; is a set of distribution features, such as skewness and kurtosis. By grouping these features into a sequence of feature blocks, the response characteristics of the sensor channel are fully described from the perspective of time and distribution. After the feature block segmentation is completed, a unique physical space identifier is assigned to each sensor channel to record the relative position information of the sensor in the array. Assuming that sensors, each of which is assigned an identifier ,in This identifier is numbered based on the actual arrangement position of the sensor, for example Indicates the physical location of the first sensor, and so on. Perform matrix concatenation on the feature block sequence and the position encoding data to construct a combined feature matrix that contains both feature information and position information. For each sensor channel, its feature block sequence and location identifier Concatenate to get the combined feature vector ; By splicing, a combined feature matrix containing sensor response characteristics and position information is constructed , which is expressed as: ; The combined feature matrix contains the time domain features, distribution features, and physical location information of all sensors in the array. The initial combined feature matrix is ​​time-segmented. The data in the combined feature matrix is ​​segmented into fixed-size time segments, and the length of each time segment is set to 100 sampling points. The original data matrix is ​​divided into multiple time segments, each of which contains the feature information of all sensor channels, and a time series block matrix is ​​obtained. The time series block matrix of each time segment It is expressed as: ; in, Contains The characteristic data of all sensor channels in a time segment. The input is fed into the temporal attention network to calculate the feature correlation score in the temporal dimension. The temporal attention network weights the features of different time segments to find the temporal features that are most important for system diagnosis. For each time segment, the attention network calculates an attention weight coefficient , which reflects the contribution of the current time segment feature to the overall feature. The calculation of the attention weight coefficient is expressed as: ; in, It is a time segment The feature importance score of is calculated by the temporal attention network. In this way, the attention network can identify and highlight the features that are most critical to the system diagnosis. Block matrix for time series Perform weighted processing to highlight key time series features. The weighted time series feature matrix is ​​expressed as: ; Through weighted processing, we get the weighted time series feature matrix , where the features of key time segments are given greater weights, so that the system can pay more attention to these important features in subsequent diagnosis. All time series features in the weighted time series feature matrix are dimensionally integrated, and the two-dimensional feature matrix is ​​compressed into a one-dimensional feature word column. Assume that the weighted feature matrix contains time segments and The data of each sensor channel is integrated according to the channel to obtain a one-dimensional feature vector : ; In this way, the original two-dimensional feature matrix is ​​compressed into a simpler one-dimensional feature sequence, retaining the important information in the time series features while reducing the complexity of the data. In order to ensure the comparability and consistency of the features, the compressed feature vector Normalization is performed. All eigenvalues ​​are uniformly mapped to the interval [0,1] to eliminate the dimension difference of the data. The formula for normalization is: ; in, is an eigenvalue in the compressed eigenvector, and are the minimum and maximum values ​​in the eigenvector respectively. The eigenvector after normalization is the spatiotemporal separation feature vector.

[0035] In a specific embodiment, the process of executing step 600 may specifically include the following steps: The spatiotemporal separation feature vectors are input forward and backward into the bidirectional long short-term memory network according to the time series, and feature learning is performed through forward propagation and backward propagation to obtain a bidirectional feature representation sequence; A feature memory queue is constructed based on the bidirectional feature representation sequence, the queue length is set to 1000 feature patterns, the queue content is updated according to the first-in-first-out principle, a dynamic feature memory library is obtained, and a cluster analysis is performed on the feature patterns stored in the dynamic feature memory library, and feature patterns with similarities higher than a fourth preset threshold are classified into the same category to obtain a normal working feature pattern set; Perform feature matching on the bidirectional feature representation sequence and the normal working feature pattern set, calculate the Euclidean distance as the feature distance deviation, obtain the feature deviation matrix, set the fault type discrimination threshold for the feature deviation matrix, mark the features with deviation values ​​greater than the fifth preset threshold as potential fault features, and obtain the fault feature marking result; Based on the fault feature labeling results, the fault modes are classified into the same category with similar deviation features to obtain the initial fault type, and the initial fault type is classified into fault levels. According to the size of the feature deviation value, the fault is divided into three levels: slight, moderate and severe to obtain the fault level classification result; The initial fault type and fault level classification results are combined to generate a fault type determination result including the fault type, fault level and fault location.

[0036] Specifically, the spatiotemporal separation feature vector is input into the BiLSTM network for feature learning. The spatiotemporal separation feature vector contains the temporal and spatial information of the sensor response features. Through forward and backward propagation, the BiLSTM network can capture the bidirectional dependencies in the time series data. Forward propagation transfers information and extracts features from the past to the present, while backward propagation processes the data in order from the future to the present. By combining these two directions, the BiLSTM network can make full use of the contextual information in the data and generate more expressive feature representations. Assume that the input spatiotemporal separation feature vector is , then the forward and backward feature representations of the BiLSTM network are recorded as: ; ; By combining the forward and backward feature representations, we get a bidirectional feature representation sequence ; Bidirectional feature representation sequence It can better describe the state and change trend of the sensor in the current time window. , construct a feature memory queue, and the length of the queue is set to 1000 feature patterns. The queue is maintained by the first-in-first-out principle, in which the earliest entering feature is removed from the queue when a new feature enters. The dynamic update mechanism can ensure that the system always maintains sensitivity to the most recent feature data, thereby quickly responding to changes in the sensor state. Over time, the feature memory queue forms a dynamic feature memory library, which stores the most recent 1000 feature patterns for continuous monitoring of the state of the sensor array. Cluster analysis is performed on the feature patterns in the dynamic feature memory library to identify normal working modes. By calculating the similarity between features, feature patterns with a similarity higher than a preset threshold (the fourth threshold) are classified into the same category to obtain a set of normal working feature patterns. Similarity is measured by cosine similarity or Euclidean distance. Assume that the feature vector is and , the calculation formula of the Euclidean distance is: ; if If the feature representation sequence is smaller than the fourth preset threshold, the two features are considered to belong to similar patterns and are classified into the same category to construct a feature pattern set containing normal working status. The bidirectional feature representation sequence is matched with the normal working feature pattern set, and the Euclidean distance between the feature of each time window and the normal pattern is calculated to obtain the feature distance deviation matrix. , with each characteristic in the normal working mode set Calculate the Euclidean distance and fill the result into the feature deviation matrix : ; in Indicates the current feature and The distance between the normal working modes. Each element of the feature deviation matrix represents the degree of difference between the feature and the normal mode. In order to determine whether the sensor has a fault, a fault type discrimination threshold (fifth threshold) is set for the feature deviation matrix. If the deviation value is greater than this preset threshold, the feature is marked as a potential fault feature. The marking result is represented by a binary flag, where 1 indicates a potential fault and 0 indicates normal. Through marking, it is preliminarily identified which sensors have characteristics that deviate from the normal working mode and have faults. Fault modes are classified based on potential fault features. Faults with similar deviation characteristics are classified into the same category to effectively identify different types of faults. Assume that there is a set of features The feature distance deviation matrix is ​​used to mark the fault features, cluster these features, and classify similar features into the same category to obtain the initial fault type. After analyzing the initial fault type, the fault is divided into three levels: slight, moderate, and severe based on the size of the feature deviation value, so as to formulate corresponding repair strategies. Set two thresholds and , used to divide the size of the deviation value: if the deviation value , it is judged as a minor fault; if , it is judged as a moderate fault; if , it is determined to be a serious fault. By dividing, the severity of each fault is described, which is convenient for subsequent fault repair and processing. The initial fault type and fault level classification results are combined to generate a fault type determination result including fault type, fault level and fault location. The fault type describes the specific mode of the fault, such as sensor response hysteresis or decreased sensitivity; the fault level describes the severity of the fault and determines the priority of repair; the fault location indicates the specific sensor channel where the problem occurs.

[0037] The above describes the intelligent fault diagnosis method of the multi-channel gas sensor in the embodiment of the present application. The following describes the intelligent fault diagnosis system of the multi-channel gas sensor in the embodiment of the present application. Figure 2 In one embodiment of the present application, an intelligent fault diagnosis system for a multi-channel gas sensor includes: A construction module 11 is used to perform random geometric analysis and response restriction area construction on the response data of n sensor channels in the multi-channel gas sensor array to obtain an effective sensor set; A screening module 12 is used to perform correlation analysis and sensitivity precision dilution measurement function screening on the response data between the sensor channels in the effective sensor set to obtain the optimal sensor combination; An extraction module 13 is used to perform periodic sampling and feature extraction on the response data of the optimal sensor combination to obtain a feature data matrix; An analysis module 14 is used to perform a deep Q network analysis on the characteristic data matrix and the sensor response change rate to generate a fault diagnosis trigger signal; The encoding module 15 is used to perform sensor channel block mapping and position encoding on the feature data matrix according to the fault diagnosis trigger signal to obtain a time-space separation feature vector; The determination module 16 is used to input the time-space separation feature vector into the bidirectional long short-term memory network to determine the fault type and output the fault type determination result.

[0038] Through the coordinated cooperation of the above-mentioned components, through random geometric analysis and response restriction area construction, a set of effective sensors with stable states is dynamically screened out, and faulty sensors are timely identified and isolated, which significantly improves the reliability of the system; the sensitivity precision dilution metric function is used for sensor optimization, which reasonably reduces the number of redundant sensors and reduces the computational complexity, while ensuring the accuracy of fault diagnosis; the fault triggering mechanism based on deep Q network and diagnostic utility function realizes the intelligent triggering of fault diagnosis and avoids the waste of computing resources caused by traditional fixed-cycle diagnosis; the temporal attention mechanism and position encoding strategy are introduced to effectively extract the spatiotemporal characteristics of the sensor and make full use of the spatial distribution information and temporal variation law of the sensor; the fault diagnosis method of the bidirectional long short-term memory network combined with the feature memory module is adopted to realize the accurate identification and classification of different types of faults, and can give the specific type, location and severity of the fault.

[0039] See also Figure 3 , Figure 3 This is a schematic block diagram of the structure of an electronic device 300 provided in an embodiment of the present application. The electronic device 300 includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected via a system bus 303, wherein the memory 302 may include a non-volatile storage medium and an internal memory.

[0040] The non-volatile storage medium can store a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor 301, the processor 301 can execute any of the above-mentioned intelligent fault diagnosis methods for multi-channel gas sensors.

[0041] The processor 301 is used to provide computing and control capabilities to support the operation of the entire electronic device 300 .

[0042] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned intelligent fault diagnosis methods for the multi-channel gas sensor.

[0043] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the electronic device 300 involved in the present application scheme. The specific electronic device 300 may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0044] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0045] It should be noted that technicians in the relevant field can clearly understand that for the convenience and simplicity of description, the specific working process of the electronic device 300 described above can refer to the corresponding process of the aforementioned intelligent fault diagnosis method of the multi-channel gas sensor, and will not be repeated here.

[0046] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by one or more processors, the one or more processors implement the intelligent fault diagnosis method for a multi-channel gas sensor provided in the embodiment of the present application.

[0047] The computer-readable storage medium may be an internal storage unit of the electronic device 300 in the aforementioned embodiment, such as a hard disk or memory of the electronic device 300. The computer-readable storage medium may also be an external storage device of the electronic device 300, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped with the electronic device 300.

[0048] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0049] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0050] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent fault diagnosis method for a multi-channel gas sensor, characterized in that: The method comprises: The response data of n sensor channels in the multi-channel gas sensor array are subjected to random geometric analysis and response restriction region construction to obtain an effective sensor set. Performing correlation analysis and sensitivity precision dilution measurement function screening on response data between sensor channels in the effective sensor set to obtain an optimal sensor combination; Performing periodic sampling and feature extraction on the response data of the optimal sensor combination to obtain a feature data matrix; Performing a deep Q network analysis on the characteristic data matrix and the sensor response change rate to generate a fault diagnosis trigger signal; According to the fault diagnosis trigger signal, the characteristic data matrix is ​​subjected to sensor channel block mapping and position encoding to obtain a time-space separation characteristic vector; The time-space separation feature vector is input into a bidirectional long short-term memory network to determine the fault type, and the fault type determination result is output.

2. The intelligent fault diagnosis method for a multi-channel gas sensor according to claim 1, characterized in that: The random geometric analysis and response restriction region construction are performed on the response data of n sensor channels in the multi-channel gas sensor array to obtain an effective sensor set, including: Collect response data of n sensor channels of the multi-channel gas sensor array, where n represents the total number of sensor channels, and obtain a response value sequence Ri (i=1, 2, ..., n), where Ri represents the response voltage value of the i-th sensor channel; The response value sequence Ri is input into the data processing unit for cumulative summation, and the sum is divided by the number of sensor channels n to obtain the response mean μ; The difference between each response value in the response value sequence Ri and the response mean μ is calculated and recorded as a difference sequence Di. The difference sequence Di is squared and then squared and divided by n-1 to obtain a response standard deviation σ; Calculate the boundary value based on the response mean μ and the response standard deviation σ, set the upper limit UL to μ+2σ, and the lower limit LL to μ-2σ, to obtain the response limit area boundary value; Compare each response value in the response value sequence Ri with the boundary value of the response limit area, and mark the response value Ri as a valid response value when it satisfies LL≤Ri≤UL, to obtain a response value marking result, wherein the response value marking result is represented by a binary number, 1 represents valid, and 0 represents invalid; Classify the n sensor channels according to the response value marking result, classify the sensor channels marked as 1 into the valid sensor set VS, and classify the sensor channels marked as 0 into the to-be-diagnosed sensor set DS; Perform response stability calculation on the sensor channels in the effective sensor set VS, and calculate the coefficient of variation CV of the response value within 10 consecutive sampling periods, wherein the coefficient of variation CV is equal to the standard deviation within the time window divided by the mean, to obtain a stability evaluation value, which is a dimensionless number; The stability evaluation value is compared with a first preset threshold value Th, wherein the first preset threshold value Th is set to 0.1, and the sensor channels whose coefficient of variation CV is lower than the first preset threshold value Th are retained in the effective sensor set VS, and the sensor channels whose coefficient of variation CV is higher than Th are transferred to the sensor set to be diagnosed DS, and the fault type of the sensor channels in the sensor set to be diagnosed is directly determined to obtain the fault determination result of the sensor set to be diagnosed.

3. The intelligent fault diagnosis method for a multi-channel gas sensor according to claim 2, characterized in that: The performing correlation analysis and sensitivity precision dilution measurement function screening on the response data between the sensor channels in the effective sensor set to obtain the optimal sensor combination includes: Calculate the Pearson correlation coefficient of the response data between any two sensor channels in the effective sensor set, where the correlation coefficient value ranges from -1 to 1, and obtain an initial correlation coefficient matrix; Normalizing the initial correlation coefficient matrix, mapping the correlation coefficient value to between 0 and 1, obtaining a standardized correlation coefficient matrix, performing a weighted sum operation on each row element in the standardized correlation coefficient matrix, and dividing the weighted sum by the number of sensor channels to obtain a weight coefficient sequence; Arranging the weight coefficient sequence into a diagonal matrix form to obtain a sensor weight matrix, and performing covariance calculation on the response data in the effective sensor set to construct a covariance matrix of the response data; Based on the sensor weight matrix and the covariance matrix, a sensitivity dilution of precision metric function SDoP value is calculated, wherein the SDoP value is the trace of the product of the weight matrix and the inverse matrix of the covariance matrix, to obtain a sensor combination evaluation index; The sensor combination evaluation index is compared with a second preset threshold value, and when the SDoP value is less than the second preset threshold value, the corresponding sensor combination is marked as a candidate combination to obtain a candidate sensor combination set; A redundancy index is calculated for each combination in the candidate sensor combination set, and a combination with the minimum redundancy and the number of sensors not less than a minimum value set by the system is selected as the optimal sensor combination.

4. The intelligent fault diagnosis method for a multi-channel gas sensor according to claim 3, characterized in that: The step of periodically sampling and extracting features from the response data of the optimal sensor combination to obtain a feature data matrix includes: The sensor channels in the optimal sensor combination are sampled for a period of 10 seconds, and a set of response signals is obtained for each sampling to obtain an original sampling data set; Performing response data normalization processing based on the maximum value Vmax and the minimum value Vmin in the original sampling data set, normalizing the data range to the interval of 0 to 1, and obtaining a normalized response data set; A data window is set for the normalized response data set, the window length is set to 100 data sampling points, and the sliding interval between adjacent windows is set to 10 data sampling points, to obtain a segmented data window set; Performing time domain statistics calculation on the data of each window in the segmented data window set, including mean, variance, peak value, and valley value, to obtain a time domain feature parameter set; Based on the time domain feature parameter set, the distribution characteristics of the data in each window are calculated, including the skewness value and the kurtosis value of the data distribution, to obtain a distribution feature parameter set, and the time domain feature parameter set and the distribution feature parameter set are feature merged to form a comprehensive feature vector of a single time window, to obtain a feature vector group; Based on the feature vector group, time series features are constructed, and feature vectors of M consecutive time windows are arranged and combined in chronological order, where M is the number of time windows, to obtain a time series feature matrix. The time series feature matrix is ​​reorganized according to the sensor channel identifier, and feature data of the same sensor channel are integrated together to obtain a feature data matrix.

5. The intelligent fault diagnosis method for a multi-channel gas sensor according to claim 4, characterized in that: The performing of a deep Q network analysis on the characteristic data matrix and the sensor response change rate to generate a fault diagnosis trigger signal includes: Performing differential calculation of adjacent time windows on the response data in the characteristic data matrix, and normalizing the data according to the sampling time interval to obtain a sensor response change rate sequence; The characteristic data matrix is ​​data-fused with the sensor response change rate to construct a system state vector, wherein the system state vector includes all statistical characteristic parameters and their change rate information, and an input state matrix of a deep Q network is obtained; Setting an action space for the input state matrix and assigning a unique identification code to each action to obtain an action space code; Constructing a deep Q network based on the input state matrix and the action space encoding, extracting and reducing the state features through convolutional layers and pooling layers to obtain state feature encoding; The state feature encoding is used to calculate the Q value of each action through a fully connected layer, and the action with the largest Q value is selected based on a greedy strategy to obtain a state-action decision value; Constructing a diagnostic utility function for the state-action decision value, performing a weighted combination of the inverse of the end-to-end diagnostic delay time and the diagnostic accuracy, and obtaining a utility function evaluation value; The utility function evaluation value is compared with a third preset threshold value, and when the utility function evaluation value is lower than the third preset threshold value, a new diagnostic process is triggered to obtain a diagnostic trigger flag, and a fault diagnosis trigger signal including a trigger time, a state number and a trigger type is generated according to the diagnostic trigger flag.

6. The intelligent fault diagnosis method for a multi-channel gas sensor according to claim 5, characterized in that: The method of performing sensor channel block mapping and position encoding on the feature data matrix according to the fault diagnosis trigger signal to obtain a time-space separation feature vector includes: Performing feature block segmentation on each sensor channel in the feature data matrix, performing grouping processing according to a time domain feature parameter set and a distribution feature parameter set to obtain a feature block sequence, and assigning a unique physical space identifier to each sensor channel based on the feature block sequence, recording relative position information in the sensor array, and obtaining position coding data; Performing matrix concatenation operation on the feature block sequence and the position coding data to construct a combined feature matrix containing both feature information and position information to obtain an initial spatiotemporal feature matrix, and dividing the data in the initial spatiotemporal feature matrix into time segments of a fixed size, each time segment being 100 sampling points long, to obtain a time series block matrix; Input the temporal block matrix into the temporal attention network, calculate the feature correlation score on the temporal dimension, obtain the temporal attention weight coefficient, and perform weighted processing on the temporal block matrix according to the temporal attention weight coefficient to highlight the key temporal features and obtain a weighted temporal feature matrix; All time series features in the weighted time series feature matrix are dimensionally integrated, the two-dimensional feature matrix is ​​compressed into a one-dimensional feature sequence to obtain a compressed feature vector, and the compressed feature vector is normalized to normalize all eigenvalues ​​to the interval [0,1] to obtain a time-space separation feature vector.

7. The intelligent fault diagnosis method for a multi-channel gas sensor according to claim 6, characterized in that: The step of inputting the time-space separation feature vector into a bidirectional long short-term memory network to determine the fault type and outputting the fault type determination result comprises: The spatiotemporal separation feature vector is input forward and backward into a bidirectional long short-term memory network according to a time series, and feature learning is performed through forward propagation and backward propagation to obtain a bidirectional feature representation sequence; A feature memory queue is constructed based on the bidirectional feature representation sequence, the queue length is set to 1000 feature patterns, the queue content is updated according to the first-in-first-out principle, a dynamic feature memory library is obtained, and a cluster analysis is performed on the feature patterns stored in the dynamic feature memory library, and feature patterns with similarities higher than a fourth preset threshold are classified into the same category to obtain a normal working feature pattern set; Perform feature matching on the bidirectional feature representation sequence and the normal working feature pattern set, calculate the Euclidean distance as the feature distance deviation, obtain a feature deviation matrix, set a fault type discrimination threshold for the feature deviation matrix, mark features with deviation values ​​greater than a fifth preset threshold as potential fault features, and obtain a fault feature marking result; Based on the fault feature labeling result, fault modes are classified into the same category with similar deviation features to obtain an initial fault type, and the initial fault type is classified into fault levels. According to the size of the feature deviation value, the fault is divided into three levels: slight, moderate and severe to obtain a fault level classification result; The initial fault type and the fault level classification result are combined to generate a fault type determination result including the fault type, fault level and fault location.

8. An intelligent fault diagnosis system for a multi-channel gas sensor, characterized in that: A method for intelligent fault diagnosis of a multi-channel gas sensor according to any one of claims 1 to 7, the system comprising: A construction module is used to perform random geometric analysis and response restriction area construction on the response data of n sensor channels in a multi-channel gas sensor array to obtain an effective sensor set; A screening module, used for performing correlation analysis and sensitivity precision dilution measurement function screening on the response data between the sensor channels in the effective sensor set to obtain the optimal sensor combination; An extraction module, used for periodically sampling and extracting features from the response data of the optimal sensor combination to obtain a feature data matrix; An analysis module, used for performing a deep Q network analysis on the characteristic data matrix and the sensor response change rate to generate a fault diagnosis trigger signal; An encoding module, used for performing sensor channel block mapping and position encoding on the characteristic data matrix according to the fault diagnosis trigger signal to obtain a time-space separation characteristic vector; The determination module is used to input the time-space separation feature vector into a bidirectional long short-term memory network to determine the fault type and output a fault type determination result.

9. An electronic device, characterized in that: The electronic device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the electronic device to execute the intelligent fault diagnosis method for the multi-channel gas sensor according to any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the intelligent fault diagnosis method for the multi-channel gas sensor according to any one of claims 1 to 7 is implemented.

Images