An airborne integrated mission processor I / O interface module fault diagnosis method and system

Abstract

The application discloses a kind of airborne integrated mission processor I / O interface module fault diagnosis method and system, it is related to avionics fault diagnosis technical field.First, part of analog quantity and digital quantity information is obtained from the I / O interface module of airborne integrated mission processor, and corresponding temperature and vibration information is obtained through temperature sensor and vibration sensor;After pre-processing and labeling, a combined model based on attention network and deep separable convolution strategy is constructed;The combined model is trained and used for real-time fault diagnosis of I / O interface module.The application triggers hardware protection for permanent fault, prompts to replace the fault module;For intermittent fault, record fault type and occurrence time, generate degradation trend report, facilitate subsequent key monitoring;The application reduces additional acquisition equipment, improves the fault diagnosis accuracy of I / O interface module of IMP, realizes accurate positioning and targeted processing of fault, and improves reliability and maintainability.

Description

Technical Field

[0001] This invention relates to the field of avionics equipment fault diagnosis technology, and in particular to a method and system for fault diagnosis of airborne integrated mission processor I / O interface module. Background Technology

[0002] In modern avionics systems, airborne integrated mission processors employ numerous and diverse I / O interface modules, such as GJB289A bus interfaces, FC bus interfaces, analog interfaces, discrete interfaces, DVI video interfaces, PAL video interfaces, and RS422 interfaces, with typically more than one of each type. These I / O interface modules play a crucial role in data transmission and command interaction between various aircraft systems. However, due to the complexity of the aviation environment and the inherent operating characteristics of I / O interface modules, they are prone to various malfunctions, severely impacting aircraft flight safety and reliability.

[0003] Traditional fault diagnosis methods typically rely on hardware redundancy or a large number of additional acquisition devices to obtain fault information. This not only increases the cost and complexity of the system, but may also lead to additional weight burden due to hardware redundancy, which is detrimental to aircraft performance improvement. In addition, traditional methods often struggle to accurately and quickly locate and diagnose faults in complex I / O interface modules, failing to meet the efficiency and accuracy requirements of modern avionics systems for fault diagnosis.

[0004] With the continuous development of sensor technology and data processing technology, data-driven fault diagnosis methods have gradually become a research hotspot. This method, through the analysis and learning of large amounts of operational data, can automatically extract fault characteristics and achieve accurate diagnosis of system faults. However, how to fully utilize data-driven methods in the fault diagnosis of the I / O interface module of airborne IMPs, reduce the need for additional data acquisition equipment, and accurately determine the fault type and location remains a problem that urgently needs to be solved.

[0005] The purpose of this invention is to provide a method and system for fault diagnosis of the I / O interface module of an airborne IMP, to solve the problems of numerous additional data acquisition devices, inaccurate fault diagnosis, and inability to accurately locate faults in existing technologies for diagnosing faults in the I / O interface module of an airborne IMP. This method and system can utilize various types of data for training to build an efficient and accurate fault diagnosis model while minimizing the number of additional data acquisition devices, enabling rapid and accurate diagnosis of I / O interface module faults and taking corresponding measures for different types of faults. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for fault diagnosis of the I / O interface module of an airborne integrated mission processor (IMP), relating to the field of avionics equipment fault diagnosis technology. First, partial analog and digital information is obtained from the I / O interface module of the IMP, and corresponding temperature and vibration information is acquired through temperature and vibration sensors. After preprocessing and annotation, a combined model based on a simplified channel attention mechanism and a depthwise separable convolution strategy is constructed. The combined model is trained and used for real-time fault diagnosis of the I / O interface module. This invention triggers hardware protection for permanent faults, prompting replacement of the faulty module; for intermittent faults, it records the fault type and occurrence time, generating a degradation trend report for convenient subsequent focused monitoring. This invention reduces the need for additional acquisition equipment, improves the accuracy of fault diagnosis of the IMP's I / O interface module, achieves precise fault location and targeted processing, and enhances reliability and maintainability.

[0007] A method and system for diagnosing faults in the I / O interface module of an airborne integrated mission processor, such as... Figure 1 As shown, it includes the following steps:

[0008] Step 1, Data Acquisition and Preprocessing: Collect multi-source feature data (KPI) and perform noise reduction and standardization processing;

[0009] Step 2, Data Labeling and Classification: Label the collected data with corresponding fault labels according to the fault conditions at the time of collection; divide the labeled data into training set and test set;

[0010] Step 3: Build a combined model (SCADSC model) based on the Simplified Channel Attention Mechanism and the Depthwise Separable Convolution strategy: The Simplified Channel Attention Mechanism strengthens key fault features by dynamically allocating channel weights; the Depthwise Separable Convolution strategy significantly reduces the computational cost while retaining the local search capability of CNN.

[0011] The SCADSC model includes an input layer, a depthwise separable convolutional module, a simplified channel attention module, a global average pooling layer, a fully connected layer, and a Softmax output layer.

[0012] Step 4, Training and optimization of the SCADSC model: Use the training set to train and optimize the SCADSC model and save the weight parameters of the trained SCADSC model to obtain the trained SCADSC model.

[0013] Step 5, test the trained SCADSC model: load weights, input the test set into the trained SCADSC model for fault classification and handling, and obtain diagnostic results; evaluate the diagnostic performance of the SCADSC model based on the diagnostic results;

[0014] Step 6: Use the trained and tested SCADSC model for fault monitoring of the I / O interface module of the airborne IMP:

[0015] The real-time collected and preprocessed data is input into the SCADSC model for fault diagnosis.

[0016] If the diagnosis result is a permanent fault, the hardware protection mechanism will be triggered, and the corresponding I / O interface module will be replaced.

[0017] If the diagnosis result is an intermittent fault, record the timestamp of the fault occurrence and generate a degradation trend report for focused monitoring.

[0018] Furthermore, in step 1, the method for collecting the KPIs is as follows:

[0019] The sampling window time is set to 5ms, and the acquisition period is set to 500ms. Each acquisition period contains 100 consecutive sampling windows. The KPI values ​​collected in each window are represented by the average KPI value within a single window. The collected data can come from simulation data fault injection during ground experiments for ground training of the diagnostic model, or from real data during actual flight to enhance the performance of the diagnostic model or achieve real-time fault monitoring.

[0020] Furthermore, in step 1, the preprocessing procedure is as follows:

[0021] Step 1.1: In each window, read some data obtained by the Airborne Integrated Mission Processor (IMP) from the I / O interface module to be diagnosed, including the ripple coefficient KPI1 and grounding impedance KPI2 of the power interface, the signal-to-noise ratio KPI3 and voltage offset rate KPI4 of the analog interface, the rise time KPI5 and fall time KPI6 of the digital interface, and the bit error rate KPI7 and response command delay KPI8 of the bus interface.

[0022] Step 1.2: Obtain vibration and temperature data of the airborne IMP during each window period using vibration and temperature sensors;

[0023] Step 1.3: For the vibration and temperature data, wavelet transform is used to remove noise interference, resulting in the processed vibration data KPI9 and temperature data KPI. 10 ;

[0024] Step 1.4: Use the Z-score standardization method to standardize all the KPIs obtained in Steps 1.1 and 1.3;

[0025] The multi-source feature data within each acquisition cycle is represented as a multi-source feature matrix. :

[0026]

[0027] Each row of data represents the processed data value of a KPI within a collection period.

[0028] Furthermore, in step 2, the data labeling and classification process is as follows:

[0029] Step 2.1: The fault types of the I / O interface module of the airborne IMP are divided into five categories: physical connection fault, vibration contact fault, electromagnetic interference fault, temperature drift fault, and power interference fault. Among them, physical connection fault is a permanent fault that cannot be recovered by itself once it occurs; the other four are intermittent faults that will recover by themselves after the environment returns to normal.

[0030] Step 2.2: Based on the fault injection method used during data acquisition, process the multi-source feature matrix. Add labels, including normal (00), physical connection failure (01), vibration contact failure (02), electromagnetic interference failure (03), temperature drift failure (04), and power interference failure (05), to obtain a feature matrix with labels. ;

[0031] Step 2.3: For the labeled data, randomly divide the training set and the test set in a 7:3 ratio.

[0032] Furthermore, the physical connection failure is caused by the I / O interface module not being properly plugged in, resulting in an open circuit or hardware damage, which in turn leads to an open circuit or short circuit in the line; physical connection failure is a permanent failure that will have serious consequences and may occur in all interfaces.

[0033] There are two methods for injecting physical connection faults:

[0034] The first method is to disconnect the I / O interface module socket from the onboard IMP;

[0035] The second method involves deliberately damaging the components of the I / O interface module, including breaking the connector pins, cutting the cable, damaging the chip of the I / O interface module, and removing the resistors and capacitors of the I / O interface module.

[0036] The vibration contact failure is caused by loose plugs or cracked solder joints due to high-frequency vibration, and is an intermittent failure. There is a failure accumulation process, and the degree of failure increases with the accumulation of usage time and the increase of environmental vibration. It can occur at all interfaces.

[0037] There are two injection methods for vibration contact faults:

[0038] The first method involves unplugging the connector between the I / O interface module and the airborne IMP by applying a radial displacement of 0.5 mm; and placing the I / O interface module and the airborne IMP connector on a vibration table and applying random frequency vibrations from 50 Hz to 2000 Hz, with the vibration acceleration randomly applied as 3 to 5 times the gravitational acceleration.

[0039] The second method involves artificially creating loose connections in some components of the I / O interface module, including worn cables, poorly soldered I / O interface module chips, resistors, and capacitors; and placing the I / O interface module and the airborne IMP connection part on a vibration table and applying random frequency vibrations of 50 to 2000 Hz, with the vibration acceleration randomly applied as 3 to 5 times the gravitational acceleration.

[0040] The electromagnetic interference fault is an intermittent fault caused by external electromagnetic interference or internal crosstalk, which occurs in high-speed digital interfaces; the degree of fault varies depending on the intensity of the electromagnetic interference, and it can usually return to normal after the interference is removed.

[0041] There are two methods for injecting electromagnetic interference faults:

[0042] The first method involves injecting 10kHz-100MHz ripple into the power lines and simultaneously injecting common-mode interference into the bus signal lines; the common-mode interference frequency is 50MHz and the voltage is 10Vpp.

[0043] The second method involves injecting pulse group interference into the interface using a pulse generator; the frequency of the pulse is 100kHz and the voltage is 2kV.

[0044] The temperature drift fault is a fault caused by changes in ambient temperature or self-heating, resulting in deviations of signal transmission accuracy, level threshold, and timing parameters from the normal range. It is temperature-dependent and reversible; it gradually worsens as the temperature rises or falls, and can be partially or completely recovered after the temperature recovers; it occurs in analog interfaces, high-precision digital interfaces, and radio frequency interfaces.

[0045] The injection method for temperature drift fault is as follows: point heating of the I / O interface module and the airborne IMP connection part is performed by using a hot air gun; the power of the hot air gun and the distance between the hot air gun and the connection part are adjusted to obtain different ambient temperature rise and fall rates;

[0046] The power interference fault is caused by voltage fluctuations, noise, and ripple, resulting in abnormal I / O interface signal transmission or function; it is an intermittent fault, and can return to normal operation after the power signal is restored; it has a significant impact on interfaces that rely on stable power supply.

[0047] There are three ways to inject power interference faults:

[0048] The first method is to use a programmable DC power supply, set the voltage step or power-off sequence, and directly connect it to the interface power supply circuit.

[0049] The second method involves generating a sinusoidal noise signal using a signal generator, which is then superimposed onto the power line via a coupling resistor; the sinusoidal noise signal has a voltage of 100mV and a frequency of 10kHz.

[0050] The third method involves using a pulse generator coupled to the power line via a capacitor to simulate instantaneous high-voltage interference.

[0051] Furthermore, in step 3, the SCADSC model includes an input layer, a depthwise separable convolutional module, a simplified channel attention module, a global average pooling layer, a fully connected layer, and a Softmax output layer;

[0052] The input dimension of the input layer of the SCADSC model is (10, 100);

[0053] The depth-separable convolution module includes two sub-modules: a first sub-module and a second sub-module.

[0054] The first submodule includes 32 depthwise convolutions with 3×3 kernels and 32 pointwise convolutions with 1×1 kernels. The output dimension of the first submodule is (32, 100).

[0055] The second submodule includes 64 depthwise convolutions with 3×3 kernels and 64 pointwise convolutions with 1×1 kernels. The output dimension of the second submodule is (64, 100).

[0056] Both submodules have a convolution kernel stride of 1, padding of 1, and ReLU activation function.

[0057] A simplified channel attention module is embedded between the two sub-modules to enhance key fault characteristics by dynamically allocating channel weights.

[0058] The simplified channel attention module includes a global average pooling layer, a fully connected layer, and a Sigmoid activation function layer. The global average pooling layer extracts the overall fault trend of the output features of the first submodule into a single statistical value, avoiding local noise interference, and simultaneously reduces the feature dimension from (32, 100) to 32 dimensions. The fully connected layer includes a dimensionality-reducing fully connected layer with 8 neurons and a dimensionality-increasing fully connected layer with 32 neurons. The dimensionality-reducing fully connected layer reduces the 32-dimensional features to 8 dimensions. The dimensionality-increasing fully connected layer maps the 8-dimensional features back to 32 dimensions. All fully connected layers adopt a linear structure, completing the dimension transformation only through linear mapping, without using an activation function. The Sigmoid activation function layer applies the Sigmoid activation function to the 32-dimensional feature vector after dimensionality increase, generating 32 weight coefficients with values ​​ranging from (0, 1).

[0059] The weighting coefficients are weighted channel-wise with the output features of the first submodule and then fed into the second submodule. The output of the second submodule is first compressed into a 64-dimensional vector by global average pooling, and then passed sequentially through a fully connected layer with 32 neurons and a Softmax output layer with 6 neurons, finally outputting the probability distribution of 6 types of faults to achieve fault classification and diagnosis. The activation function of the fully connected layer is ReLU activation with a Dropout probability of 0.2. Figure 2 As shown.

[0060] Furthermore, in step 4, the training and optimization methods for the SCADSC model are as follows:

[0061] The Adam optimizer is used with an initial learning rate of 0.001. Cosine annealing is used to adjust the learning rate, which is reduced to 0.8 times the current rate every 5 iterations to prevent local optima. The number of iterations (Epoch) is set to 50, and the objective function for training the SCADSC model is the cross-entropy loss function.

[0062] A fault diagnosis system for an airborne integrated mission processor I / O interface module is provided, the fault diagnosis system being used to execute a fault diagnosis method for an airborne integrated mission processor I / O interface module.

[0063] The beneficial effects of this invention are as follows:

[0064] This invention collects information using only the existing IMP interface data and a small number of temperature and vibration sensors (common equipment in aircraft), without the need for additional dedicated acquisition devices, thus reducing additional equipment and system complexity; the SCADSC model focuses on key features through an attention mechanism, and depthwise separable convolution efficiently extracts local fault modes, thereby improving diagnostic accuracy;

[0065] This invention triggers hardware protection for permanent faults, prompting the replacement of the faulty module. For intermittent faults, it records the fault type and occurrence time, generating a degradation trend report for easy monitoring later. It distinguishes between permanent and intermittent faults, enabling immediate replacement of permanent faults and early warning of intermittent fault trends, reducing unnecessary downtime for maintenance, improving aircraft availability, enabling targeted fault handling, and enhancing reliability and maintainability. The model has low computational requirements and can be deployed in airborne embedded systems to meet real-time diagnostic needs, demonstrating strong engineering practicality. Attached Figure Description

[0066] Figure 1 This is an overall flowchart of the present invention;

[0067] Figure 2 This is a schematic diagram of the SCADSC model;

[0068] Figure 3 This is a schematic diagram of an airborne integrated mission processor I / O interface module fault diagnosis system according to an embodiment of the present invention. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. A method and system for diagnosing faults in the I / O interface module of an airborne integrated mission processor, such as… Figure 1 As shown, it includes the following steps:

[0070] Step 1, Data Acquisition and Preprocessing: Read the multi-source characteristic data (KPIs) obtained by the airborne IMP from the I / O interface module during fault injection, including the ripple coefficient KPI1 and grounding impedance KPI2 of the power interface, the signal-to-noise ratio KPI3 and voltage offset KPI4 of the analog interface, the rise time KPI5 and fall time KPI6 of the digital interface, and the bit error rate KPI7 and response command delay KPI8 of the bus interface. Also, acquire vibration and temperature data from the airborne IMP within the same time period using vibration and temperature sensors. When reading the multi-source characteristic data, a 500ms acquisition cycle is used, with each cycle containing 100 consecutive sampling windows. The KPI values ​​acquired within each window are represented by the average value of that KPI within that single window. For each fault injection period (including fault-free injection), 1000 cycles of data are collected. For vibration and temperature data, wavelet transform is used to remove noise interference, resulting in processed vibration data KPI9 and temperature data KPIs. 10 All KPIs are standardized using the Z-score standardization method; the multi-source feature data within each acquisition period consists of a multi-source feature matrix. This indicates that each row of data represents the processed data value of a KPI within one collection period:

[0071]

[0072] The fault types of the airborne IMP's I / O interface module are categorized into five main types: physical connection faults, vibration contact faults, electromagnetic interference faults, temperature drift faults, and power supply interference faults. Physical connection faults are permanent and cannot recover automatically once they occur; the other four types are intermittent faults that recover automatically once the environment returns to normal. Fault injection methods during data acquisition primarily target the FC bus interface, and the specific injection methods are as follows:

[0073] (1) Physical connection failure:

[0074] Disconnect the FC bus interface module socket from the onboard IMP;

[0075] The process involved breaking some pins of the FC bus interface module, cutting some cables, damaging some chips, and removing some resistors and capacitors.

[0076] (2) Vibration contact fault:

[0077] Disconnect the plugs at the FC bus interface module and the airborne IMP connection (apply 0.5 mm radial displacement of the plug), and place the FC bus interface module and the airborne IMP connection on a vibration table and apply random frequency vibration of 50 to 2000 Hz, with the vibration acceleration randomly applied as 3 to 5 times the gravitational acceleration;

[0078] Wear down some cables of the FC bus interface module and poorly solder some chips, resistors and capacitors of the FC bus interface module. At the same time, place the FC bus interface module and the airborne IMP connection part on a vibration table and apply random frequency vibration of 50 to 2000 Hz, and apply vibration acceleration of 3 to 5 times the acceleration of gravity.

[0079] (3) Electromagnetic interference fault:

[0080] Simultaneously, a 10kHz-100MHz ripple is injected into the power line, and common-mode interference is injected into the FC bus interface. The frequency of the common-mode interference is 50MHz, and the voltage is 10Vpp.

[0081] A pulse group interference is injected into the FC bus interface through a pulse generator. The frequency of the pulse is 100kHz and the voltage is 2kV.

[0082] (4) Temperature drift fault:

[0083] By using a hot air gun to heat the I / O interface module and the airborne IMP connection part at specific points, different environmental heating and cooling rates can be obtained by adjusting the power of the hot air gun and the distance between it and the connection part.

[0084] (5) Power supply interference fault:

[0085] Use a programmable DC power supply to set the voltage step;

[0086] A sinusoidal noise signal is generated by a signal generator and superimposed on the power line through a coupling resistor; the voltage of the sinusoidal noise signal is 100mV and the frequency is 10kHz.

[0087] Step 2, Data Labeling and Classification: Based on the fault injection method used during data collection in Step 1, the matrix is ​​labeled and classified. Add labels, including normal (00), physical connection failure (01), vibration contact failure (02), electromagnetic interference failure (03), temperature drift failure (04), and power interference failure (05), to obtain a feature matrix with labels. The labeled data was randomly divided into training and test sets in a 7:3 ratio. The distribution of the labeled sample data is shown in Table 1.

[0088] surface Sample data distribution

[0089] Data types Label Number of training set samples Number of test set samples normal 00 700 300 Physical connection failure 01 700 300 Vibration contact fault 02 700 300 Electromagnetic interference fault 03 700 300 Temperature drift fault 04 700 300 Power interference fault 05 700 300

[0090] Step 3, Build the SCADSC model: Establish a combined model (SCADSC model) based on the Simplified Channel Attention Mechanism and the Depthwise Separable Convolution strategy. The Simplified Channel Attention Mechanism strengthens key fault features by dynamically allocating channel weights; the Depthwise Separable Convolution strategy significantly reduces computational cost while retaining the local search capabilities of CNNs. Figure 2As shown, the SCADSC model includes an input layer, a depthwise separable convolutional module, a simplified channel attention module, a global average pooling layer, a fully connected layer, and a softmax output layer. The input layer has an input dimension of (10, 100). The depthwise separable convolutional module consists of two sub-modules: the first sub-module includes 32 depthwise convolutions with 3×3 kernels and 32 pointwise convolutions with 1×1 kernels, with an output dimension of (32, 100); the second sub-module includes 64 depthwise convolutions with 3×3 kernels and 64 pointwise convolutions with 1×1 kernels, with an output dimension of (64, 100). Both sub-modules have convolutional kernel strides of 1, padding of 1, and ReLU activation functions. A simplified channel attention module is embedded between the two sub-modules to enhance key fault features through dynamic allocation of channel weights. The simplified channel attention module includes a global average pooling layer, a fully connected layer, and a Sigmoid activation function layer. The global average pooling layer of the simplified channel attention module extracts the overall fault trend from the output features of the first sub-module of the depthwise separable convolutional module into a single statistical value, avoiding local noise interference, and reducing the feature dimension from (32, 100) to 32 dimensions. The fully connected layer of the simplified channel attention module includes two fully connected layers: dimensionality reduction and dimensionality enhancement. The dimensionality reduction part uses a fully connected layer with 8 neurons to enhance the dimensionality of the features. The 2D features are reduced to 8D; the dimensionality reduction part is then passed through a 32-neuron fully connected layer to map the 8D features back to 32D; the Sigmoid activation function layer of the simplified channel attention module applies the Sigmoid activation function to the 32D feature vector after dimensionality reduction, generating 32 weight coefficients with values ​​in the range (0,1); these are then weighted channel-wise with the output features of the first submodule and passed to the second submodule; the output of the second submodule is compressed into a 64D vector by global average pooling, and then passed through a 32-neuron fully connected layer (ReLU activation, Dropout probability 0.2) and an output layer (Softmax activation) with 6 neurons, finally outputting the probability distribution of 6 types of faults to achieve fault classification and diagnosis.

[0091] Step 4: Train and optimize the SCADSC model: Use the training dataset to train and optimize the SCADSC model and save the weight parameters of the trained SCADSC model to obtain the trained SCADSC model; During training, the Adam optimizer is used, the initial learning rate is set to 0.001, and cosine annealing is used to adjust the learning rate. Every 5 iterations, the learning rate is decayed to 0.8 times the current rate to prevent local optima. The number of iterations (Epoch) is set to 50, and the objective function for training the SCADSC model is the cross-entropy loss function.

[0092] Step 5, Fault Classification and Handling: Load weights, input the test set into the trained SCADSC model, and obtain diagnostic results. Evaluate the diagnostic performance of the SCADSC model based on the results.

[0093] Step 6: Use the trained SCADSC model for fault monitoring of the I / O interface module of the airborne IMP. For example... Figure 3 As shown, the airborne integrated mission processor I / O interface module fault diagnosis system completes data acquisition and processing through steps 1 and 2. Then, in step 3, a SCADSC model is built, and the model is trained on the ground using either simulation experimental data or real data (step 4). The trained SCADSC model can then be used for real-time fault diagnosis of the actual aircraft. After collecting and processing the actual aircraft flight data, it is input into the trained SCADSC model to complete real-time monitoring of the I / O interface module. Based on the detection results, appropriate actions are taken. If the diagnosis result is a permanent fault, a hardware protection mechanism is triggered, and the corresponding I / O interface module is replaced. If the diagnosis result is various types of intermittent faults, the specific type of fault and the timestamp of the fault occurrence are recorded, and a degradation trend report is generated for focused monitoring.

Claims

1. A method for diagnosing faults in the I / O interface module of an airborne integrated mission processor, characterized in that, The fault diagnosis method includes the following steps: Step 1, Data Acquisition and Preprocessing: Collect multi-source feature data KPIs and perform noise reduction and standardization processing; Step 2, Data Labeling and Classification: Label the collected data with corresponding fault labels according to the fault conditions at the time of collection; divide the labeled data into training set and test set; Step 3: Build the SCADSC model based on the simplified channel attention mechanism and the depthwise separable convolution strategy: The simplified channel attention mechanism strengthens key fault features by dynamically allocating channel weights; the depthwise separable convolution strategy significantly reduces the amount of computation while retaining the local search capability of CNN. The SCADSC model includes an input layer, a depthwise separable convolutional module, a simplified channel attention module, a global average pooling layer, a fully connected layer, and a Softmax output layer. The SCADSC model includes an input layer, a depthwise separable convolutional module, a simplified channel attention module, a global average pooling layer, a fully connected layer, and a Softmax output layer. The input dimension of the input layer of the SCADSC model is (10, 100); The depth-separable convolution module includes two sub-modules: a first sub-module and a second sub-module. The first submodule includes 32 depthwise convolutions with 3×3 kernels and 32 pointwise convolutions with 1×1 kernels. The output dimension of the first submodule is (32, 100). The second submodule includes 64 depthwise convolutions with 3×3 kernels and 64 pointwise convolutions with 1×1 kernels. The output dimension of the second submodule is (64, 100). Both submodules have a convolution kernel stride of 1, padding of 1, and ReLU activation function. A simplified channel attention module is embedded between the two sub-modules to enhance key fault characteristics by dynamically allocating channel weights. The simplified channel attention module includes a global average pooling layer, a fully connected layer, and a Sigmoid activation function layer. The global average pooling layer extracts the overall fault trend of the output features of the first submodule into a single statistical value, avoiding local noise interference, and simultaneously reduces the feature dimension from (32, 100) to 32 dimensions. The fully connected layer includes a dimensionality-reducing fully connected layer with 8 neurons and a dimensionality-increasing fully connected layer with 32 neurons. The dimensionality-reducing fully connected layer reduces the 32-dimensional features to 8 dimensions. The dimensionality-increasing fully connected layer maps the 8-dimensional features back to 32 dimensions. All fully connected layers adopt a linear structure, completing the dimension transformation only through linear mapping, without using an activation function. The Sigmoid activation function layer applies the Sigmoid activation function to the 32-dimensional feature vector after dimensionality increase, generating 32 weight coefficients with values ​​ranging from (0, 1). The SCADSC model includes an input layer, a depthwise separable convolutional module, a simplified channel attention module, a global average pooling layer, a fully connected layer, and a Softmax output layer. The input dimension of the input layer of the SCADSC model is (10, 100); The depth-separable convolution module includes two sub-modules: a first sub-module and a second sub-module. The first submodule includes 32 depthwise convolutions with 3×3 kernels and 32 pointwise convolutions with 1×1 kernels. The output dimension of the first submodule is (32, 100). The second submodule includes 64 depthwise convolutions with 3×3 kernels and 64 pointwise convolutions with 1×1 kernels. The output dimension of the second submodule is (64, 100). Both submodules have a convolution kernel stride of 1, padding of 1, and ReLU activation function. A simplified channel attention module is embedded between the two sub-modules to enhance key fault characteristics by dynamically allocating channel weights. The simplified channel attention module includes a global average pooling layer, a fully connected layer, and a Sigmoid activation function layer. The global average pooling layer extracts the overall fault trend of the output features of the first submodule into a single statistical value, avoiding local noise interference, and simultaneously reduces the feature dimension from (32, 100) to 32 dimensions. The fully connected layer includes a dimensionality-reducing fully connected layer with 8 neurons and a dimensionality-increasing fully connected layer with 32 neurons. The dimensionality-reducing fully connected layer reduces the 32-dimensional features to 8 dimensions. The dimensionality-increasing fully connected layer maps the 8-dimensional features back to 32 dimensions. All fully connected layers adopt a linear structure, completing the dimension transformation only through linear mapping, without using an activation function. The Sigmoid activation function layer applies the Sigmoid activation function to the 32-dimensional feature vector after dimensionality increase, generating 32 weight coefficients with values ​​ranging from (0, 1). Step 4, Training and optimization of the SCADSC model: Use the training set to train and optimize the SCADSC model and save the weight parameters of the trained SCADSC model to obtain the trained SCADSC model. Step 5, test the trained SCADSC model: load weights, input the test set into the trained SCADSC model for fault classification and handling, and obtain diagnostic results; evaluate the diagnostic performance of the SCADSC model based on the diagnostic results; Step 6: Use the trained and tested SCADSC model for fault monitoring of the I / O interface module of the airborne IMP: The real-time collected and preprocessed data is input into the SCADSC model for fault diagnosis. If the diagnosis result is a permanent fault, the hardware protection mechanism will be triggered, and the corresponding I / O interface module will be replaced. If the diagnosis result is an intermittent fault, record the timestamp of the fault occurrence and generate a degradation trend report for focused monitoring.

2. The method for fault diagnosis of an airborne integrated mission processor I / O interface module according to claim 1, characterized in that, In step 1, the KPI collection method is as follows: The sampling window time is set to 5ms, the acquisition period is set to 500ms, and each acquisition period contains 100 consecutive sampling windows. The KPI values ​​collected in each window are represented by the average KPI value within a single window.

3. The method for fault diagnosis of an airborne integrated mission processor I / O interface module according to claim 1, characterized in that, In step 1, the preprocessing process is as follows: Step 1.1: Read some data obtained by the airborne IMP from the I / O interface module to be diagnosed in each window, including the ripple coefficient KPI1 and grounding impedance KPI2 of the power interface, the signal-to-noise ratio KPI3 and voltage offset rate KPI4 of the analog interface, the rise time KPI5 and fall time KPI6 of the digital interface, and the bit error rate KPI7 and response command delay KPI8 of the bus interface. Step 1.2: Obtain vibration and temperature data of the airborne IMP during each window period using vibration and temperature sensors; Step 1.3: For the vibration and temperature data, wavelet transform is used to remove noise interference, resulting in the processed vibration data KPI9 and temperature data KPI. 10 ; Step 1.4: Use the Z-score standardization method to standardize all the KPIs obtained in Steps 1.1 and 1.3; The multi-source feature data within each acquisition cycle is represented as a multi-source feature matrix. : ； Each row of data represents the processed data value of a KPI within a collection period.

4. The method for fault diagnosis of an airborne integrated mission processor I / O interface module according to claim 1, characterized in that, In step 2, the data labeling and classification process is as follows: Step 2.1: The fault types of the I / O interface module of the airborne IMP are divided into five categories: physical connection fault, vibration contact fault, electromagnetic interference fault, temperature drift fault, and power interference fault. Among them, physical connection fault is a permanent fault that cannot be recovered by itself once it occurs; the other four are intermittent faults that will recover by themselves after the environment returns to normal. Step 2.2: Based on the fault injection method used during data acquisition, process the multi-source feature matrix. Add labels, including normal, physical connection fault, vibration contact fault, electromagnetic interference fault, temperature drift fault, and power interference fault, to obtain a feature matrix with labels. ; Step 2.3: For the labeled data, randomly divide the training set and the test set in a 7:3 ratio.

5. A method for diagnosing faults in an airborne integrated mission processor I / O interface module according to claim 4, characterized in that, The physical connection failure is caused by the I / O interface module not being properly plugged in, resulting in an open circuit or hardware damage, which in turn leads to an open circuit or short circuit in the line; physical connection failure is a permanent failure. There are two methods for injecting physical connection faults: The first method is to disconnect the I / O interface module socket from the onboard IMP; The second method involves deliberately damaging the components of the I / O interface module, including breaking the connector pins, cutting the cable, damaging the chip of the I / O interface module, and removing the resistors and capacitors of the I / O interface module. The vibration contact failure is caused by loose plugs or cracked solder joints due to high-frequency vibration, and is an intermittent failure. There are two injection methods for vibration contact faults: The first method involves unplugging the connector between the I / O interface module and the airborne IMP by applying a radial displacement of 0.5 mm; and placing the I / O interface module and the airborne IMP connector on a vibration table and applying random frequency vibrations from 50 Hz to 2000 Hz, with the vibration acceleration randomly applied as 3 to 5 times the gravitational acceleration. The second method involves artificially creating loose connections in some components of the I / O interface module, including worn cables, poorly soldered I / O interface module chips, resistors, and capacitors; and placing the I / O interface module and the airborne IMP connection part on a vibration table and applying random frequency vibrations of 50 to 2000 Hz, with the vibration acceleration randomly applied as 3 to 5 times the gravitational acceleration. The electromagnetic interference fault is an intermittent fault caused by external electromagnetic interference or internal crosstalk, which occurs in high-speed digital interfaces. There are two methods for injecting electromagnetic interference faults: The first method involves injecting 10kHz-100MHz ripple into the power lines and simultaneously injecting common-mode interference into the bus signal lines; the common-mode interference has a frequency of 50MHz and a voltage of 10Vpp. The second method involves injecting pulse group interference into the interface using a pulse generator. The frequency of the pulse is 100kHz and the voltage is 2kV. The temperature drift fault is a fault caused by changes in ambient temperature or self-heating, resulting in deviations of signal transmission accuracy, level threshold, and timing parameters from the normal range. It is temperature-dependent and reversible. The injection method for temperature drift fault is as follows: point heating of the I / O interface module and the airborne IMP connection part is performed by using a hot air gun; the power of the hot air gun and the distance between the hot air gun and the connection part are adjusted to obtain different ambient temperature rise and fall rates; The power interference fault is caused by voltage fluctuations, noise, and ripple, resulting in abnormal I / O interface signal transmission or function. It is an intermittent fault, and can return to normal operation after the power signal is restored; it has a significant impact on interfaces that rely on stable power supply. There are three ways to inject power interference faults: The first method is to use a programmable DC power supply, set the voltage step or power-off sequence, and directly connect it to the interface power supply circuit. The second method involves generating a sinusoidal noise signal using a signal generator, which is then superimposed onto the power line via a coupling resistor; the sinusoidal noise signal has a voltage of 100mV and a frequency of 10kHz. The third method involves using a pulse generator coupled to the power line via a capacitor to simulate instantaneous high-voltage interference.

6. The method for fault diagnosis of an airborne integrated mission processor I / O interface module according to claim 1, characterized in that, In step 3, the weight coefficients are weighted channel-wise with the output features of the first submodule and then fed into the second submodule. The output of the second submodule is first compressed into a 64-dimensional vector by global average pooling, and then passed through a fully connected layer with 32 neurons and a Softmax output layer with 6 neurons in sequence to finally output the probability distribution of 6 types of faults, thereby realizing fault classification and diagnosis. The activation function of the fully connected layer is ReLU activation with a Dropout probability of 0.

2.

7. A method for diagnosing faults in an airborne integrated mission processor I / O interface module according to claim 1, characterized in that, In step 4, the training and optimization methods for the SCADSC model are as follows: The Adam optimizer was used, with an initial learning rate of 0.

001. Cosine annealing was used to adjust the learning rate, which was decayed to 0.8 times the current rate every 5 iterations. The number of iterations was set to 50, and the objective function for training the SCADSC model was the cross-entropy loss function.

8. A fault diagnosis system for an airborne integrated mission processor I / O interface module, wherein the fault diagnosis system adopts the fault diagnosis method for an airborne integrated mission processor I / O interface module as described in any one of claims 1 to 7.

Images