Fault diagnosis method for filter circuit of airborne integrated task processor

Abstract

The invention discloses an airborne integrated task processor filter circuit fault diagnosis method, and belongs to the technical field of airborne electronic system fault diagnosis. Firstly, an airborne IMP fault classification model is constructed; secondly, acquiring fault data from an IMP filter circuit according to an airborne IMP fault classification model and preprocessing the fault data; and then training a multi-cavity convolution LSTM model, extracting local spatial features through a multi-cavity convolution layer, fusing the spatial-temporal features by a BiLSTM layer, and realizing classification of normal, intermittent and permanent faults by using a Softmax classification layer. The problem that the false alarm rate is high due to the fact that intermittent faults are ignored in a traditional method is solved, and the diagnosis accuracy is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of airborne electronic system fault diagnosis, and particularly relates to a fault diagnosis method for a filter circuit of an airborne integrated mission processor. BACKGROUND

[0002] As a core component of modern avionics architecture, the integrated mission processor (IMP) is prone to be affected by faults in harsh environments such as high vibration and complex electromagnetic interference, which brings challenges to fault prediction and health management. The existing fault diagnosis method for the airborne integrated mission processor (IMP) mainly adopts a two-state fault model (normal / permanent fault), which ignores the transient characteristics of intermittent faults, resulting in an increased risk of false positives.

[0003] Intermittent faults occur 10-30 times more frequently than permanent faults in IMPs, caused by poor contact, component degradation or environmental mutations, and have the characteristics of transience, randomness and non-repetition, which are extremely difficult to detect in routine maintenance. The traditional method does not consider the intermittent fault state, which is prone to false positives for slight intermittent faults when dealing with IMPs with tightly coupled architecture, reducing the accuracy of diagnosis, increasing maintenance costs, and even affecting system stability. In addition, the natural collection cost of intermittent fault data of the airborne IMP system is extremely high, and the real fault samples are scarce due to flight task limitations, resulting in insufficient training data for existing models.

[0004] In the existing data-driven fault diagnosis method, traditional shallow learning needs to extract features manually, and deep learning methods such as CNN, DBN and RNN have problems such as lack of global context, gradient explosion / disappearance, etc., making it difficult to effectively fuse the spatiotemporal features of the IMP circuit degradation dynamics. Therefore, there is an urgent need for a method that can accurately depict the IMP fault degradation process and achieve efficient diagnosis. SUMMARY

[0005] To solve the problems of low accuracy and difficult data collection of the existing intermittent fault diagnosis of the airborne IMP, the present application provides a fault diagnosis method for a filter circuit of an airborne integrated mission processor, which belongs to the technical field of airborne electronic system fault diagnosis. First, a fault classification model for the airborne IMP is constructed; second, fault data is obtained from the IMP filter circuit according to the fault classification model for the airborne IMP and preprocessed; then a multiple-hole convolution LSTM model is trained, local spatial features are extracted through the multiple-hole convolution layer, temporal and spatial features are fused through the BiLSTM layer, and the classification of normal, intermittent and permanent faults is realized through the Softmax classification layer. The present application solves the problem of high false positive rate caused by the traditional method ignoring intermittent faults, and significantly improves the accuracy of diagnosis.

[0006] A fault diagnosis method for a filter circuit of an airborne integrated mission processor, as shown in Figure 1 the figure, comprising the following steps:

[0007] Step 1, constructing an airborne IMP fault classification model;

[0008] The airborne IMP fault classification model is a three-state fault degradation model;

[0009] The airborne IMP fault classification model classifies three states of the airborne IMP filter circuit, including normal working state NF, intermittent fault state IF and permanent fault state PF;

[0010] The intermittent fault state includes three different severity states of slight intermittent fault IF1, moderate intermittent fault IF2 and severe intermittent fault IF3;

[0011] When fault injection, the different states of the airborne IMP filter circuit are determined by the different switching periods of the high-speed switch programmed by the STM32 single-chip microcomputer;

[0012] Step 2, obtaining fault data;

[0013] For the airborne IMP filter circuit, five states of the circuit are obtained by controlling the different states of resistors, capacitors and operational amplifiers through high-speed switches, including normal working state NF, slight intermittent fault IF1, moderate intermittent fault IF2, severe intermittent fault IF3 and permanent fault state PF;

[0014] By applying a sinusoidal excitation signal to the airborne IMP filter circuit, the voltage signal at the output end of the airborne IMP filter circuit is collected using an oscilloscope as fault data;

[0015] According to the state of the airborne IMP filter circuit during fault data collection, add labels corresponding to the five states, namely NF label, IF1 label, IF2 label, IF3 label and PF label; according to the label classification of the five states, store to the signal storage device;

[0016] Step 3, data preprocessing; the data preprocessing is to sequentially perform downsampling, normalization and sample segmentation;

[0017] First, for the fault data obtained in step 2, after downsampling and normalization, the continuous signal is converted to fixed-length sample data through sliding window;

[0018] Then, the fixed-length sample data is randomly divided into training set and test set by sample segmentation method in the proportion of 7:3; the training set and the test set form a data set;

[0019] The downsampling is used to retain the 0-300Hz frequency band, which is the key frequency band, to reduce computational redundancy;

[0020] The normalization eliminates dimensional influence by Z-score standardization;

[0021] Step 4, a multiple hollow convolution LSTM model is built;

[0022] The multiple hollow convolution LSTM model is an MCLSTM model, and the MCLSTM model comprises a multiple hollow convolution layer, a bidirectional long short-term memory network layer (i.e. a BiLSTM layer) and a Softmax classification layer.

[0023] Step 5, training and optimization of the MCLSTM model;

[0024] The MCLSTM model is trained and optimized by using the training set, and the MCLSTM model weight parameters after training and optimization are saved to obtain a trained MCLSTM model;

[0025] Step 6, result output and fault monitoring;

[0026] The parameter weight of the MCLSTM model is loaded, the test set is input into the trained MCLSTM model to obtain a diagnosis result, the diagnosis performance of the MCLSTM model is evaluated according to the diagnosis accuracy on the test set, and the tested MCLSTM model is used for fault monitoring of the airborne IMP filter circuit.

[0027] Further, in step 1, the airborne IMP fault classification model is:

[0028] ;

[0029] Wherein, represents an actual measurement signal of the airborne IMP filter circuit; represents a signal when the airborne IMP filter circuit works normally; represents a fault signal; 、 and are functions of time t, and satisfies: ;

[0030] Wherein, is a unit step signal; and respectively represent the start and end times of the i-th intermittent fault, and satisfy , represents a fault condition, which is determined by the generation mechanism of the intermittent fault; as shown in Figure 2 ;

[0031] According to different values of , the state of the airborne IMP filter circuit is divided into NF, IF and PF; when Constantly = 0, that is, Constantly = 0, that is, When the on-board IMP filter circuit is in NF;

[0032] When The on-board IMP filter circuit is in IF; the intermittent appearance is that Constantly ≠ 0, but in the case of = 0; the more the number i of fault signals, the longer the fault duration The more serious the fault is; according to the different severity of the fault, IF is subdivided into three levels of IF1, IF2 and IF3; IF1 refers to the proportion of the time of the existence of the fault signal ≠ 0 time) is greater than 0 but less than or equal to 20%, IF2 refers to the proportion of the existence of the fault signal is greater than 20% but less than or equal to 50%, and IF3 refers to the proportion of the existence of the fault signal is greater than 50% but less than 100%;

[0033] When Constantly exists and does not appear = 0, the on-board IMP filter circuit is in PF, indicating that the circuit element has been permanently damaged and cannot work normally.

[0034] Further, in step 2, the method for obtaining fault data is as follows:

[0035] The on-board IMP successively includes a signal source module, a to-be-tested on-board IMP filter circuit, an STM32 single-chip microcomputer and a data acquisition module; the signal source module includes a direct current power supply and a sine wave generator; the data acquisition module includes an oscilloscope and a signal storage device; as Figure 3 The direct current power supply respectively provides ±12V and 5V direct current power supply for the to-be-tested on-board IMP filter circuit and the STM32 single-chip microcomputer, and the sine wave generator generates a 5V, 25kHz sine wave Vsin as an input signal of the to-be-tested on-board IMP filter circuit; the STM32 single-chip microcomputer realizes the function of high-speed switching through programming control, and controls the to-be-tested on-board IMP filter circuit to be in different states through different switching periods of the high-speed switch; in each state of the to-be-tested on-board IMP filter circuit, the high-precision oscilloscope in the data acquisition module acquires the voltage signal at the output end of the to-be-tested on-board IMP filter circuit as fault data, and transmits the fault data into the signal storage device for storage.

[0036] Further, the to-be-tested on-board IMP filter circuit is a four-operational-amplifier double-second-order high-pass filter circuit; as Figure 4As shown, the airborne IMP four operational amplifier double second-order high-pass filter circuit includes UA741 operational amplifier U, resistor R and capacitor C; the UA741 operational amplifier includes operational amplifiers U1, U2, U3 and U4, the resistor R includes resistors R1, R2, R3, R4, R5, R6, R7, R8, R9 and R10, and the capacitor C includes capacitors C1 and C2;

[0037] The sinusoidal wave Vsin as the input signal enters the input end of the measured airborne IMP filter circuit and is divided into two branches:

[0038] The sinusoidal wave Vsin of the first branch is transmitted to the inverting input end of U1 through R3; R1 and C1 are connected in parallel, one end of which is connected to the inverting input end of U1, and the other end is connected to the output end of U1; the output end of U1 is connected to R9 and R4, R9 is connected to the inverting input end of U4, and R4 is connected to the inverting input end of U2; one end of C2 is connected to the inverting input end of U2, and the other end is connected to the output end of U2; the output end of U2 is transmitted to the inverting input end of U3 through R5; the output end of U3 is divided into three paths, which are respectively fed back to the inverting input end of U3 through R6, to the inverting input end of U1 through R2, and to the inverting input end of U4 through R7;

[0039] The sinusoidal wave Vsin of the second branch is directly connected to the inverting input end of U4 after being transmitted through R10; one end of R8 is connected to the inverting input end of U4, and the other end is connected to the output end of U4; the output end of U4 is the final output signal Vout;

[0040] The non-inverting input ends of U1, U2, U3 and U4 are all connected to the ground, and U1, U2, U3 and U4 are all powered by +12V and -12V dual DC power supply;

[0041] The high-speed switch controls the state of the airborne IMP filter circuit by controlling the access state of the elements, which include R1, C1 and U1:

[0042] When the high-speed switch directly connects a high-resistance resistor (100kΩ) in series with R1, R1 appears an open circuit fault; when the high-speed switch does not connect any resistor element in series with R1, the circuit is in a normal state;

[0043] Similarly, the high-speed switch controls the open circuit fault and normal state of C1 and U1 by controlling R1.

[0044] Further, when using an oscilloscope to sample and collect the voltage signal at the output end of the airborne IMP filter circuit, the sampling period of each sampling is 4ms, the sampling frequency is 250kHZ, and the sample length is 1000; each sampling period is recorded as a data acquisition period, and the control equation of the high-speed switch in one data acquisition period is:

[0045] ​

[0046] in, Represents a data collection cycle; and Respectively represent Duration of the secondary fault period and the non-fault period; Indicates the number of occurrences of the failure period;

[0047] During the non-fault period, the high-speed switch allows the component to be normally connected to the circuit;

[0048] The fault period includes the fault active state and fault failure status ; and The number of switching times between two states during a fault period is expressed as ;

[0049] During the fault period, the high-speed switch switches at different frequencies, and the fault is activated. When a high-resistance resistor is connected in series to the circuit, the component will fail in an open circuit; When the component is normally connected to the circuit, the circuit is normal;

[0050] By programming the STM32 microcontroller to control the different switching cycles of the high-speed switch, we can obtain 、 、 duration and duration, thereby controlling the different fault states of R1, C1 and U1;

[0051] When the high-speed switch keeps the components connected to the circuit normally throughout the data acquisition period, the circuit is in NF;

[0052] When the high-speed switch keeps the component in open-circuit fault throughout the data acquisition period, the circuit is in PF;

[0053] When the high-speed switch is switched within a data acquisition cycle to ensure that the total fault activation state time (i.e., the fault signal existence time) within a data acquisition cycle is at different proportions, the circuit is in an intermittent fault state to varying degrees.

[0054] Furthermore, in step 4, the multiple dilated convolution layers use 9 convolution kernels; the size of the convolution kernel is 3×1, the number is 64, and the expansion rate is The exponential increase is as follows, where j is the convolutional layer number; the first layer has a dilation rate of 1 and a receptive field of 3 data points; the last layer has a dilation rate of 256 and a receptive field of 1023 data points; the output of each layer is activated by ReLU and batch normalized;

[0055] The BiLSTM layer is two layers of BiLSTM, each layer containing 128 units.

[0056] The Softmax classification layer is an output layer.

[0057] Further, in step 5, the MCLSTM model training method is as follows:

[0058] The fault data in the training set is input into the MCLSTM model, and the weight parameters of the MCLSTM model are obtained after iterative training;

[0059] In the training, the Adam optimizer is used, the learning rate is set to 0.001, the batch size is 32, and the iteration number is 50 times;

[0060] The MCLSTM model parameters are updated through back propagation, a regularization method is introduced to prevent overfitting, and the learning rate is dynamically adjusted;

[0061] The objective function for training the MCLSTM model is the cross-entropy loss function , which is used to calculate the difference between the predicted fault state and the real fault state, and the cross-entropy loss function The expression is:

[0062] ;

[0063] Wherein, N is the sample number; C is the number of fault state categories; k is the sample index, c is the fault category index, is the real label; is the predicted probability;

[0064] The cross-entropy loss function is used to measure the difference between the predicted probability of the MCLSTM model and the real label, guide the MCLSTM model to continuously adjust the parameters, and improve the prediction accuracy.

[0065] An electronic device comprising: one or more processors; memory; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs are configured to execute the above-mentioned airborne integrated task processor filter circuit fault diagnosis method.

[0066] A computer readable storage medium stores program code, the program code can be called by a processor to execute the above-mentioned airborne integrated task processor filter circuit fault diagnosis method.

[0067] The beneficial effects of the present application are as follows:

[0068] Solve the problem of data scarcity: the intermittent fault data scarcity problem of airborne IMP is effectively solved through the three-state fault model and programmable high-speed switch control strategy; by simulating intermittent faults of different severity and degradation process to permanent fault in the laboratory environment, a large number of representative samples covering normal, intermittent fault and permanent fault states can be efficiently generated, providing sufficient data support for model training and breaking the training bottleneck caused by the difficulty of obtaining real fault data in traditional methods;

[0069] Improve diagnostic accuracy: the MCLSTM model can efficiently fuse the spatio-temporal features of IMP fault signals through the synergistic effect of multi-dilation convolution layer and bidirectional LSTM layer; the multi-dilation convolution layer expands the receptive field with exponentially increasing dilation rate, extracts multi-scale local spatial features and captures subtle changes in fault signals; the bidirectional LSTM layer fully utilizes the bidirectional time dependence of the sequence to understand the time evolution law of the fault signal; experiments show that the average diagnostic accuracy of the present application is 97.85%, which is 13.34%, 6.77% and 4.23% higher than that of BP, BiLSTM and CNN methods respectively, significantly improving the diagnostic accuracy of three-state faults;

[0070] Reduce false positive rate: traditional two-state models ignore intermittent fault states and are prone to false positives for slight intermittent faults; the three-state model of the present application accurately depicts the intermediate transition state of IMP from normal to permanent fault, can accurately identify intermittent faults of different severity, effectively reduces the false positive rate of diagnosis, provides reliable technical support for predictive maintenance of avionics systems, reduces unnecessary maintenance costs and improves system stability;

[0071] Strong generalization ability: the present application is not only suitable for airborne IMP filter circuit, but also can be applied to other similar electronic circuit systems; by reasonably adjusting the model parameters and fault injection strategy, it can adapt to the fault diagnosis needs of different types of components and different working environments, has strong generalization ability and engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is the overall flowchart of the present application;

[0073] Figure 2 is a three-state fault degradation model;

[0074] Figure 3 is a fault acquisition circuit schematic diagram;

[0075] Figure 4 is a schematic diagram of the four operational amplifier double second-order high-pass filter circuit of the airborne IMP;

[0076] Figure 5 is the confusion matrix of the MCLSTM model diagnosis result. DETAILED DESCRIPTION

[0077] The application will be further described in detail below in combination with the drawings and examples.

[0078] A method for diagnosing faults of an airborne integrated mission processor filter circuit, as shown in the figure, comprises the following steps: Figure 1

[0079] Step 1: Construct an airborne IMP fault classification model.

[0080] The airborne IMP fault classification model is a three-state fault degradation model.

[0081] The airborne IMP fault classification model classifies three states of the airborne IMP filter circuit, including a normal working state (NF), an intermittent fault state (IF) and a permanent fault state (PF).

[0082] The intermittent fault state includes three different severity states of a slight intermittent fault (IF1), a moderate intermittent fault (IF2) and a severe intermittent fault (IF3).

[0083] Different states of the airborne IMP filter circuit are determined by different switching periods of high-speed switches programmed and controlled by an STM32 single-chip microcomputer.

[0084] The airborne IMP fault classification model is:

[0085] ;

[0086] wherein, represents an actual measured signal of the airborne IMP filter circuit; represents a signal when the airborne IMP filter circuit is in normal working state; represents a fault signal; 、 and are functions of time t, and satisfy: ;

[0087] wherein, is a unit step signal; and respectively represent start and end times of the i-th intermittent fault, and satisfy , represents a fault condition, which is determined by the generation mechanism of the intermittent fault; as shown in the figure; Figure 2

[0088] According to different values of , the state of the airborne IMP filter circuit is divided into NF, IF and PF; when ​​Constant is 0, that is Constant is equal to When the airborne IMP filter circuit is in NF;

[0089] When Intermittent in time function, the airborne IMP filter circuit is in IF; the intermittent appearance is that Not constant at 0, but in the case of =0; the more the number i of fault signals appears, the longer the fault duration The more serious the fault is; according to the different severity of the fault, IF is subdivided into three levels of IF1, IF2 and IF3; IF1 refers to the proportion of fault signal existing time (time ≠0) is greater than 0 but less than or equal to 20%, IF2 refers to the proportion of fault signal existing time is greater than 20% but less than or equal to 50%, IF3 refers to the proportion of fault signal existing time is greater than 50% but less than 100%;

[0090] When Constantly exists and does not appear =0, the airborne IMP filter circuit is in PF, indicating that the circuit elements have been permanently damaged and cannot work normally.

[0091] Step 2, get fault data;

[0092] For the airborne IMP filter circuit, five states of the circuit are obtained by high-speed switch control of resistance, capacitance and operational amplifier in different states, and the five states are NF, IF1, IF2, IF3 and PF;

[0093] By applying a sinusoidal excitation signal to the airborne IMP filter circuit, the voltage signal at the output end of the airborne IMP filter circuit is collected using an oscilloscope as fault data;

[0094] According to the state of the airborne IMP filter circuit when collecting fault data, add labels corresponding to the five states, namely NF label, IF1 label, IF2 label, IF3 label and PF label; according to the labels of the five states, store to the signal storage device;

[0095] The method for obtaining fault data is as follows:

[0096] As Figure 3As shown, the airborne IMP in turn includes a signal source module, a to-be-tested airborne IMP filter circuit, an STM32 single-chip microcomputer, and a data acquisition module; the signal source module includes a direct current power supply and a sine wave generator; the data acquisition module includes an oscilloscope and a signal storage device; the direct current power supply provides ±12V and 5V direct current power supplies for the to-be-tested airborne IMP filter circuit and the STM32 single-chip microcomputer, respectively, and the sine wave generator generates a 5V, 25kHz sine wave Vsin as an input signal of the to-be-tested airborne IMP filter circuit; the STM32 single-chip microcomputer realizes the function of high-speed switching through programming control, controls the to-be-tested airborne IMP filter circuit to be in different states through different switching periods of the high-speed switch; within each state of the to-be-tested airborne IMP filter circuit, a high-precision oscilloscope in the data acquisition module acquires a voltage signal at an output end of the to-be-tested airborne IMP filter circuit as fault data, and transmits the fault data into the signal storage device for storage;

[0097] As shown in Figure 4 , the to-be-tested airborne IMP filter circuit is a four-operational-amplifier double-second-order high-pass filter circuit of the airborne IMP; the four-operational-amplifier double-second-order high-pass filter circuit of the airborne IMP includes UA741 operational amplifiers U, resistors R, and capacitors C; the UA741 operational amplifiers include operational amplifiers U1, U2, U3, and U4, the resistors R include resistors R1, R2, R3, R4, R5, R6, R7, R8, R9, and R10, and the capacitors C include capacitors C1 and C2;

[0098] After the sine wave Vsin as the input signal enters an input end of the to-be-tested airborne IMP filter circuit, the sine wave Vsin is divided into two branches:

[0099] The sine wave Vsin in the first branch is transmitted to the inverting input end of U1 through R3; R1 and C1 are connected in parallel, one end of which is connected to the inverting input end of U1, and the other end of which is connected to the output end of U1; the output end of U1 is connected to R9 and R4, respectively, R9 is connected to the inverting input end of U4, and R4 is connected to the inverting input end of U2; one end of C2 is connected to the inverting input end of U2, and the other end of C2 is connected to the output end of U2; the output end of U2 is transmitted to the inverting input end of U3 through R5; the output end of U3 is divided into three branches, which are respectively fed back to the inverting input end of U3 through R6, fed back to the inverting input end of U1 through R2, and transmitted to the inverting input end of U4 through R7;

[0100] The sine wave Vsin in the second branch is transmitted through R10 and then directly connected to the inverting input end of U4; one end of R8 is connected to the inverting input end of U4, and the other end of R8 is connected to the output end of U4; the output end of U4 is the final output signal Vout;

[0101] The non-inverting input ends of U1, U2, U3, and U4 are all connected to the ground, and U1, U2, U3, and U4 are all powered by +12V and -12V dual direct current power supplies;

[0102] High-speed switch controls the state of onboard IMP filter circuit by controlling the access state of control elements, which include R1, C1 and U1:

[0103] When high-speed switch connects high resistance (100kΩ) with R1 in series, R1 appears open circuit fault; when high-speed switch connects no resistance element with R1 in series, the circuit is in normal state;

[0104] Similarly, high-speed switch controls the open circuit fault and normal state of C1 and U1 by controlling R1;

[0105] When using oscilloscope to sample and collect the voltage signal of the output end of onboard IMP filter circuit, the sampling period is 4ms each time, the sampling frequency is 250kHZ, and the sample length is 1000; record each sampling period as a data acquisition period, and the control equation of high-speed switch in one data acquisition period is:

[0106]

[0107] Wherein, represents a data acquisition period; and respectively represent the duration of the th fault period and the non-fault period; represents the number of fault periods;

[0108] In the non-fault period, high-speed switch makes the element normally access the circuit;

[0109] The fault period includes fault activation state and fault failure state ; and The number of switches between the two states in a fault period is represented as ;

[0110] In the fault period, high-speed switch switches at different frequencies, fault activation state connects high resistance into the circuit, and the element open circuit fault; fault failure state connects the element normally into the circuit, and the circuit is normal;

[0111] By programming STM32 single-chip microcomputer to control the different switching periods of high-speed switch, the duration of , , and the duration of are obtained, so as to control the different fault states of R1, C1 and U1;

[0112] ​When the high-speed switch keeps the element normally connected to the circuit throughout the entire data acquisition period, the circuit is in NF;

[0113] When the high-speed switch keeps the element in open-circuit failure throughout the entire data acquisition period, the circuit is in PF;

[0114] When the high-speed switch switches at different frequencies in one data acquisition period, ensuring that the total failure activation state time (i.e., the time when the failure signal exists) in one data acquisition period is in different proportions, the circuit is in intermittent failure state of different degrees.

[0115] The control parameters of intermittent failures of different levels are shown in Table 1. In Table 1, represents the number of failure periods, and represent the time of failure activation state and failure inactivation state, respectively, N s represents the number of switches in the failure activation state in one failure period. Using a random number generation algorithm, the time point and injection point of intermittent failure occurrence are determined according to the interval shown in Table 1, and the high-speed switch is controlled by the STM32 single-chip microcomputer to quickly change in different states, randomly generating intermittent failures of different intensities and durations.

[0116] Table 1. Parameter settings for intermittent failures of different levels

[0117] Failure level

[0012] N p (times)]]>

[0012] N s (times)]]> T a / μm T b / μm Level 1 [1,2] [1,4] [1,5] [1,3] Level 2 [3,6] [5,10] [5,10] [1,6] Level 3 [7,12] [11,15] [10,20] [1,10]

[0118] Step 3, data preprocessing;

[0119] The data preprocessing includes downsampling, normalization and sample segmentation in sequence;

[0120] First, for the failure data obtained in step 2, after downsampling and normalization, the continuous signal is converted to fixed-length sample data through a sliding window;

[0121] Then, the fixed-length sample data is randomly divided into training set and test set in the proportion of 7:3 through sample segmentation; the training set and the test set form a data set;

[0122] The downsampling is used to retain the key frequency band of 0-300Hz and reduce computational redundancy;

[0123] The normalization adopts Z-score standardization to eliminate the influence of dimension;

[0124] The training set includes 700 samples and corresponding labels, and the test set includes 300 samples and corresponding labels;

[0125] The data set contains 13,000 labeled samples, as shown in Table 2.

[0126] Table 2 Sample distribution under different conditions

[0127] Failure label Failure code Training sample Test sample F0 NF 700 300 F1 IF1-R1 700 300 F2 IF2-R1 700 300 F3 IF3-R1 700 300 F4 PF-R1 700 300 F5 IF1-U1 700 300 F6 IF2-U1 700 300 F7 IF3-U1 700 300 F8 PF-U1 700 300 F9 IF1-C1 700 300 F10 IF2-C1 700 300 F11 IF3-C1 700 300 F12 PF-C1 700 300

[0128] Step 4, build a multiple hollow convolution LSTM model;

[0129] The multiple hollow convolution LSTM model is an MCLSTM model, which includes a multiple hollow convolution layer, a bidirectional long short-term memory network layer (BiLSTM layer) and a Softmax classification layer.

[0130] The multiple hollow convolution layer uses a 9-layer convolution kernel; the size of the convolution kernel is 3x1, the number is 64, and the expansion rate is exponentially increased, j is the convolution layer number; the expansion rate of the first layer is 1, and the receptive field is 3 data points; the expansion rate of the last layer is 256, and the receptive field is 1023 data points; each layer output is activated by ReLU and batch normalized;

[0131] The BiLSTM layer is two layers of BiLSTM, each containing 128 units.

[0132] The Softmax classification layer is the output layer.

[0133] The forward LSTM learns historical dependencies, and the backward LSTM captures future features, and the combined output forms a 256-dimensional time series feature vector; finally, the 13-class fault state probability (including normal state and 12 fault states) is output through the Softmax layer.

[0134] Step 5, training and optimization of the MCLSTM model;

[0135] The training set is used to train and optimize the MCLSTM model, and the weight parameters of the trained and optimized MCLSTM model are saved to obtain the trained MCLSTM model;

[0136] The MCLSTM model training method is as follows:

[0137] The fault data in the training set is input into the MCLSTM model, and the weight parameters of the MCLSTM model are obtained after iterative training;

[0138] In the training, the Adam optimizer is used, the learning rate is set to 0.001, the batch size is 32, and the iteration number is 50 times.

[0139] ​The MCLSTM model parameters are updated through backpropagation, and the regularization method is introduced to prevent overfitting, so as to avoid the situation where the MCLSTM model performs well on the training set but poorly on the test set. At the same time, the learning rate is dynamically adjusted. When the rate of decrease of the training loss function slows down, the learning rate is appropriately reduced to improve the convergence accuracy of the model. The objective function of training the MCLSTM model is the cross entropy loss function. , cross entropy loss function expression:

[0140] ;

[0141] Where N is the number of samples; C is the number of fault status categories; k is the sample index, and c is the fault category index. is the true label; The cross entropy loss function is used to measure the difference between the model's predicted probability and the true label, guiding the model to continuously adjust parameters to improve the accuracy of the prediction.

[0142] Step 6: Result output and fault monitoring;

[0143] Load the parameter weights of the MCLSTM model; input the test set into the trained MCLSTM model to obtain the diagnostic results; evaluate the diagnostic performance of the MCLSTM model based on the diagnostic accuracy on the test set; and use the tested MCLSTM model for fault monitoring of the airborne IMP filter circuit.

[0144] The experimental results are shown in Table 3:

[0145] Table 3 Diagnostic results and performance of the circuit under test

[0146] Failure label Failure code Precision Recall F0 NF 1.000 1.000 F1 IF1-R1 0.872 0.977 F2 IF2-R1 0.977 0.993 F3 IF3-R1 1.000 0.997 F4 PF-R1 1.000 1.000 F5 IF1-U1 0.975 0.923 F6 IF2-U1 0.990 0.977 F7 IF3-U1 1.000 0.993 F8 PF-U1 0.997 1.000 F9 IF1-C1 0.965 0.910 F10 IF2-C1 0.976 0.957 F11 IF3-C1 0.980 0.993 F12 PF-C1 1.000 1.000

[0147] The average accuracy of the MCLSTM model reached 97.85%. Confusion matrix of the diagnosis results Figure 5 As shown in Figure 2, except for the intermittent fault F1, the accuracy of the MCLSTM model on other fault types has reached more than 96.5%. The confusion matrix of the diagnosis results is shown in Figure 2. Figure 5 As shown in the figure, except for the IF1 intermittent faults F5 and F9, which have a maximum misclassification of 0.07 on the F1 fault type, the proposed method can achieve an accuracy of more than 98% on other fault types.

Claims

1. An airborne integrated mission processor filter circuit fault diagnosis method, characterized by, Comprising the following steps: Step 1, build an airborne IMP fault classification model; The airborne IMP fault classification model classifies three states of the airborne IMP filter circuit, including the normal working state NF, the intermittent fault state IF and the permanent fault state PF; The intermittent fault state includes three different severity states of slight intermittent fault IF1, medium intermittent fault IF2 and severe intermittent fault IF3; During fault injection, the different states of the airborne IMP filter circuit are determined by the different switching periods of the high-speed switch programmed by the STM32 single-chip microcomputer; Step 2, obtain fault data; For the airborne IMP filter circuit, five states are obtained by controlling the different states of the resistance, capacitance and operational amplifier through the high-speed switch, including the normal working state NF, the slight intermittent fault IF1, the medium intermittent fault IF2, the severe intermittent fault IF3 and the permanent fault state PF; By applying a sinusoidal excitation signal to the airborne IMP filter circuit, the voltage signal at the output end of the airborne IMP filter circuit is collected using an oscilloscope as fault data; According to the state of the airborne IMP filter circuit during fault data collection, add labels corresponding to the five states, namely NF label, IF1 label, IF2 label, IF3 label and PF label; store the five state labels to the signal storage device according to the five state labels; Step 3, data preprocessing; the data preprocessing is to sequentially perform downsampling, normalization and sample segmentation; First, for the fault data obtained in step 2, after downsampling and normalization, the continuous signal is converted to fixed-length sample data through sliding window; Then, the fixed-length sample data is randomly divided into training set and test set in the proportion of 7:3 through sample segmentation; the training set and the test set form a data set; The downsampling retains the 0-300Hz frequency band; the 0-300Hz frequency band is the key frequency band; The normalization adopts Z-score standardization; Step 4, build a multiple hollow convolution LSTM model; The multiple hollow convolution LSTM model is the MCLSTM model, which includes a multiple hollow convolution layer, a BiLSTM layer and a Softmax classification layer; Step 5, training and optimization of the MCLSTM model; The training set is used to train and optimize the MCLSTM model, and the weight parameters of the trained and optimized MCLSTM model are saved to obtain the trained MCLSTM model; Step 6, result output and fault monitoring; Load the parameter weight of the MCLSTM model; Input the test set into the trained MCLSTM model to obtain the diagnosis result; evaluate the diagnosis performance of the MCLSTM model according to the diagnosis accuracy on the test set; use the tested MCLSTM model for fault monitoring of the airborne IMP filter circuit.

2. The method of claim 1, wherein the method further comprises: In step 1, the airborne IMP fault classification model is: ； wherein represents the actual measured signal of the on-board IMP filter circuit; represents the signal of the on-board IMP filter circuit when it is working properly; represents the fault signal; , and are functions of time t, and satisfy: ; wherein is a unit step signal; and denote the start and end time of the i-th intermittent fault, respectively, and satisfy , denotes the fault condition, which is determined by the generation mechanism of the intermittent fault; According to different values of the IMP filter circuit state is divided into NF, IF and PF; when is always 0, that is is equal to , the IMP filter circuit is in NF; When The airborne IMP filter circuit is in IF when it intermittently appears as a function of time, i.e. Not equal to 0, but in the case of =0; the more the number of fault signals i, the longer the fault duration The more serious the fault is; according to the different severity of the fault, IF is subdivided into three levels of IF1, IF2 and IF3; IF1 refers to the fault signal existing time ratio greater than 0 but less than or equal to 20%, IF2 refers to the fault signal existing time ratio greater than 20% but less than or equal to 50%, and IF3 refers to the fault signal existing time ratio greater than 50% but less than 100%. When Constantly present and not appearing = 0, the on-board IMP filter circuit is in PF, indicating that the circuit elements have been permanently damaged and cannot work properly.

3. The method of claim 1, wherein the method further comprises: In step 2, the method for obtaining fault data is as follows: The airborne IMP successively comprises a signal source module, a to-be-tested airborne IMP filter circuit, an STM32 single-chip microcomputer and a data acquisition module; the signal source module comprises a direct current power supply and a sine wave generator; the data acquisition module comprises an oscilloscope and a signal storage device; The direct current power supply respectively provides the to-be-tested airborne IMP filter circuit and the STM32 single-chip microcomputer with ±12V and 5V direct current power supplies, the sine wave generator generates a sine wave Vsin as an input signal of the to-be-tested airborne IMP filter circuit, the voltage of the sine wave Vsin is 5V, and the frequency is 25kHz; the STM32 single-chip microcomputer realizes the function of high-speed switching through programming control, controls the to-be-tested airborne IMP filter circuit to be in different states through different switching periods of the high-speed switch; in each state of the to-be-tested airborne IMP filter circuit, a high-precision oscilloscope in the data acquisition module acquires a voltage signal at an output end of the to-be-tested airborne IMP filter circuit as fault data, and transmits the fault data into the signal storage device for storage.

4. The method of claim 3, wherein the method further comprises: The to-be-tested airborne IMP filter circuit is a four-operational-amplifier double-second-order high-pass filter circuit of the airborne IMP; the four-operational-amplifier double-second-order high-pass filter circuit of the airborne IMP comprises UA741 operational amplifiers U, resistors R and capacitors C; the UA741 operational amplifiers comprise operational amplifiers U1, U2, U3 and U4, the resistors R comprise resistors R1, R2, R3, R4, R5, R6, R7, R8, R9 and R10, and the capacitors C comprise capacitors C1 and C2; After the sine wave Vsin as the input signal enters an input end of the to-be-tested airborne IMP filter circuit, the sine wave Vsin is divided into two branches: The first branch of the sine wave Vsin is transmitted to the inverting input end of U1 through R3; R1 and C1 are connected in parallel, one end of which is connected to the inverting input end of U1, and the other end is connected to the output end of U1; the output end of U1 is connected to R9 and R4, R9 is connected to the inverting input end of U4, and R4 is connected to the inverting input end of U2; one end of C2 is connected to the inverting input end of U2, and the other end is connected to the output end of U2; the output end of U2 is transmitted to the inverting input end of U3 through R5; the output end of U3 is divided into three branches, which are respectively fed back to the inverting input end of U3 through R6, fed back to the inverting input end of U1 through R2, and transmitted to the inverting input end of U4 through R7; The second branch of the sine wave Vsin is transmitted through R10, and then directly connected to the inverting input end of U4; one end of R8 is connected to the inverting input end of U4, and the other end is connected to the output end of U4; the output end of U4 is the final output signal Vout; The non-inverting input ends of U1, U2, U3 and U4 are all connected to the ground, and U1, U2, U3 and U4 are all powered by +12V and -12V double direct current power supplies; The high-speed switch controls the state of the airborne IMP filter circuit by controlling the connection state of the elements, which comprises R1, C1 and U1: When the high-speed switch directly connects the high-resistance resistor 100kΩ and R1 in series, R1 appears an open circuit fault; when the high-speed switch does not connect any resistor element and R1 in series, the circuit is in a normal state; Similarly, the high-speed switch controls the open circuit fault and the normal state of C1 and U1 by controlling R1.

5. The method of claim 3, wherein the method further comprises: When the voltage signal at the output end of the airborne IMP filter circuit is sampled and collected using an oscilloscope, each sampling period is regarded as a data collection period, and the control equation of the high-speed switch in one data collection period is: ​ ； wherein, represents a data collection period; and represent the duration of the first failure period and the non-failure period, respectively; represents the number of occurrences of failure periods; In the non-fault period, the high-speed switch makes the element normally access the circuit; The failure period comprises a failure active state and a failure inactive state ; and The number of switches between the two states within a failure period is denoted by ; In the fault period, the high-speed switch is switched at different frequencies, and the fault activation state When the high resistance resistor is connected in series in the circuit, the element open circuit fault; Fault failure state When the element is normally connected to the circuit, the circuit is normal; By programming control of different switching periods of high-speed switches by STM32 single-chip microcomputer, durations of , , , and durations of are obtained, thereby controlling different fault states of R1, C1 and U1; When the high-speed switch keeps the element normally access the circuit in the whole data acquisition period, the circuit is in NF; When the high-speed switch keeps the element in open circuit fault in the whole data acquisition period, the circuit is in PF; When the high-speed switch switches in a data acquisition period, ensuring that the total fault activation state time in a data acquisition period is in different proportion, the circuit is in intermittent fault state of different degrees.

6. The method of claim 1, wherein: In step 4, the multiple dilated convolution layers use 9 convolution kernels; the size of the convolution kernel is 3×1, the number is 64, and the expansion rate is The exponential increase is as follows, where j is the convolutional layer number; the first layer has a dilation rate of 1 and a receptive field of 3 data points; the last layer has a dilation rate of 256 and a receptive field of 1023 data points; the output of each layer is activated by ReLU and batch normalized; The BiLSTM layer is two layers of BiLSTM, each layer containing 128 units; The Softmax classification layer is an output layer.

7. The method of claim 1, wherein: In step 5, the MCLSTM model training method is as follows: Input the fault data in the training set into the MCLSTM model, and obtain the weight parameters of the MCLSTM model after iterative training; Update the MCLSTM model parameters by back propagation, introduce the regularization method to prevent overfitting, and dynamically adjust the learning rate; The objective function for training the MCLSTM model is a cross-entropy loss function for calculating the difference between the predicted failure state and the real failure state, the cross-entropy loss function is expressed as: ； where N is the number of samples; C is the number of failure status categories; k is the sample index, and c is the failure category index, is the true label; is the predicted probability.

8. An electronic device, comprising: It comprises: One or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured to execute the method of any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program code, and the program code can be called and executed by the processor to execute the method of any one of claims 1 to 7.

Images