An intelligent diagnosis method and system based on an IMP self-test module

Abstract

The application relates to an intelligent diagnosis method and system based on an IMP self-test module, and relates to the field of electronic equipment health management. The intelligent diagnosis system is a voltage-current dual-mode diagnosis system, sequentially comprising a multi-source signal acquisition module, a parallel computing module, an evidence fusion module and a diagnosis report generation module; first, multi-source signals are acquired through the multi-source signal acquisition module, wherein the multi-source signals comprise voltage signals and current signals; then, the multi-source signals are introduced into the parallel computing module for processing to obtain a feature vector; then, the feature vector is input into the evidence fusion module as an evidence source to perform decision fusion through evidence theory; finally, a diagnosis result is output through the diagnosis report generation module. The application has the characteristics of multi-source complementation, voltage-current heterogeneous sensing covers the impedance characteristics and energy transmission characteristics of a circuit, solves the problem of insufficient representation of a single signal source, has a feature decoupling mechanism, an inflation convolution is good at spatial abnormal distribution, a Bi-LSTM focuses on time sequence evolution rules, and feature interference is avoided.

Description

Technical Field

[0001] This invention relates to the field of health management of electronic devices, and specifically to an intelligent diagnostic method and system based on an IMP self-test module. Background Technology

[0002] As a core component of modern avionics architecture, the Integrated Mission Processor (IMP) is highly susceptible to failure when operating in harsh environments such as high vibration and complex electromagnetic interference, posing challenges to fault prediction and health management.

[0003] The ever-increasing complexity of modern electronic systems presents significant challenges to fault diagnosis. Flight control systems in aerospace equipment typically contain over 200 functional submodules and more than 300 sensors, forming a tightly coupled monitoring network. This complex structure amplifies the risk of localized fault propagation: when a transient anomaly occurs in an operational amplifier circuit, the built-in test system may trigger a false alarm mechanism, leading to a chain reaction of redundant module switching and functional degradation. Statistical data shows that the false alarm rate of a certain type of satellite power control system reached as high as 5.2% during on-orbit operation, with 87% of the false alarms originating from misjudgments of intermittent faults in analog circuits.

[0004] Intermittent failures are essentially the uncertain responses of electronic components under dynamic environmental stress. Physically, this manifests as: periodic switching on / off of solder joints due to microcracks during temperature cycling; fluctuations in equivalent series resistance caused by electrolyte drying in aluminum electrolytic capacitors; and contact impedance jumps in thin-film resistors due to mechanical vibration. These failures have three core characteristics: transient anomalies on the order of microseconds to milliseconds in time; drift of local node parameters in space; and a cumulative deterioration effect in causality. Experimental measurements show that under temperature cycling conditions from -40℃ to 125℃, the intermittent failure rate of a certain power amplifier output stage is 15 times that of a permanent failure; and under 5g vibration acceleration, the capacitance value of the filter capacitor can fluctuate by ±30% of its nominal value.

[0005] Existing diagnostic technologies have three limitations: single-point monitoring schemes are limited by spatial resolution, and the diagnostic accuracy for multi-node coupled faults is less than 75%; spatial feature extraction methods based on convolutional neural networks are slow to respond to transient signals, and the false negative rate exceeds 40% in intermittent fault detection; although bidirectional long short-term memory networks can capture time-series dependencies, they are not sensitive to spatial distribution features such as resistance parameter drift. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides an intelligent diagnostic method and system based on an IMP self-test module, relating to the field of electronic device health management. The intelligent diagnostic system is a voltage-current dual-modal diagnostic system, comprising a multi-source signal acquisition module, a parallel computing module, an evidence fusion module, and a diagnostic report generation module. First, the multi-source signal acquisition module acquires multi-source signals, including voltage and current signals. Then, the multi-source signals are imported into the parallel computing module for processing to obtain feature vectors. These feature vectors are then used as evidence sources and input to the evidence fusion module for decision fusion based on evidence theory. Finally, the diagnostic report generation module outputs the diagnostic results. This invention features multi-source complementarity, with voltage-current heterogeneous sensing covering circuit impedance characteristics and energy transfer features, solving the problem of insufficient single-signal-source representation. It also features a feature decoupling mechanism, with dilated convolution specifically targeting spatial anomalies and Bi-LSTM focusing on temporal evolution patterns to avoid feature interference.

[0007] A smart diagnostic method based on the IMP self-test module, such as Figure 1 As shown, it includes the following steps:

[0008] Step 1: Create a fault model classification;

[0009] Intermittent faults in the Integrated Mission Processor (IMP) self-test module are categorized into circuit component resistor R faults, capacitor C faults, and operational amplifier U faults; fault states are categorized into normal (NF), intermittent fault (IF), and permanent fault (PF).

[0010] The intermittent faults are classified into three levels according to their severity: minor intermittent faults (IF1), moderate intermittent faults (IF2), and severe intermittent faults (IF3).

[0011] Get the set of fault types The set of fault types Includes 13 types of failure modes;

[0012] ;

[0013] Among them, RIF1 indicates a minor intermittent resistor fault, RIF2 indicates a moderate intermittent resistor fault, RIF3 indicates a severe intermittent resistor fault, RPF indicates a permanent resistor fault, CIF1 indicates a minor intermittent capacitor fault, CIF2 indicates a moderate intermittent capacitor fault, CIF3 indicates a severe intermittent capacitor fault, CPF indicates a permanent capacitor fault, UIF1 indicates a minor intermittent op-amp fault, UIF2 indicates a moderate intermittent op-amp fault, UIF3 indicates a severe intermittent op-amp fault, and UPF indicates a permanent op-amp fault.

[0014] Step 2: Deploy voltage monitoring points;

[0015] The voltage monitoring point includes the input terminal U of the IMP self-test module circuit. in The circuit consists of the first-stage amplifier output terminal U1, the bandpass node BP, the key node of the RC network C1, the key node of the RC network R0, and the circuit output terminal U4.

[0016] Step 3: The multi-source signal acquisition module uses a voltage monitoring sensor to acquire the voltage signal V(t) and the voltage drop Vr(t) across the sampling resistor;

[0017] Step 4: Calculate the current signal I(t);

[0018] In the charging and discharging path C of energy storage capacitor C1 path With load terminal L path power load R P A precision sampling resistor is connected in series, and a compensation capacitor C2 is set simultaneously, with C2 connected in parallel across the sampling resistor; the current signal I(t) is obtained according to the current conversion formula I(t)=Vr(t) / Rr, where Rr=0.1Ω±0.1%;

[0019] Step 5: Use an anti-aliasing filter to acquire the voltage signal V(t) and generate a time series dataset;

[0020] Step 6: The voltage signal V(t), the resistance voltage drop Vr(t), and the current signal I(t) are downsampled as raw signal data to obtain the corresponding slice signals.

[0021] Step 7: Input the voltage signal V(t) slice signal into the parallel computing module, and use a nine-layer dilated convolutional network to extract spatial features, obtaining a 128-dimensional feature vector F. mc ;

[0022] Step 8: Similarly, the sliced ​​current signal I(t) is input into the parallel computing module and bidirectional long short-term memory network to obtain the 128-dimensional feature vector F. bi ;

[0023] Step 9: The evidence fusion module uses evidence theory to analyze the 128-dimensional feature vector F. mc and 128-dimensional eigenvector F bi After decision fusion, the results are input into the diagnostic report generation module, which then outputs the fault diagnosis result A*.

[0024] Furthermore, in step 3, the voltage signal V(t) and the voltage drop across the sampling resistor Vr(t) are synchronously acquired at a sampling rate of 750kHz.

[0025] Furthermore, in step 4, the accuracy level of the sampling resistor is 0.02.

[0026] Furthermore, in step 5, the passband cutoff frequency of the anti-aliasing filter is 375kHz, and the continuous acquisition time is 3 seconds.

[0027] Furthermore, in step 6, the downsampling settings are as follows:

[0028] The sampling rate was first reduced from 750kHz to 250kHz before data acquisition, resulting in 600 sets of samples.

[0029] During data acquisition, the acquisition lasted for 3 seconds, with a signal period of approximately 4ms; each sample group had a length of 1024 points.

[0030] Furthermore, in step 7, the layer parameters of the nine-layer dilated convolutional network include the receptive field and the dilation factor;

[0031] Employing a zero-fill strategy to maintain output size stability, receptive field The extended recurrence formula is:

[0032] (1);

[0033] (2);

[0034] Where l represents the number of dilated convolution layers; Indicates the output of the current layer; represents the input of the current layer; k represents the size of the convolution kernel; the receptive field of the first layer is R1=3; the ninth layer is 1023 points; the size of the output feature matrix of the nine-layer dilated convolutional network is decreased by formulas (1) and (2).

[0035] Furthermore, in step 7, the 128-dimensional feature vector F mc The acquisition process is as follows:

[0036] First, the voltage signal V(t) slice signal is passed through a nine-layer dilated convolutional network to output a 4×64 feature matrix;

[0037] Each layer of the dilated convolutional network is configured with 64 3×1 convolutional kernels with a fixed stride of 1 and an inflation factor d. l =2^(l-1) doubling layer by layer, where l represents the number of dilated convolutional layers;

[0038] Then, the 4×64 feature matrix is ​​flattened and compressed into a 128-dimensional feature vector F by a 128-node fully connected layer. mc .

[0039] Furthermore, in step 8, the 128-dimensional feature vector F bi The acquisition process is as follows:

[0040] First, the 1024×1 current signal I(t) slice signal is processed by a 128-processing-unit bidirectional long short-term memory network to synchronously calculate the forward propagation state and the backward propagation state, outputting the hidden state matrix, which is then reduced to a 1024×128 spatiotemporal feature matrix H through a 128-node fully connected layer.

[0041] (3);

[0042] Among them, h j Let represent the feature vector at time step j, where j ranges from [1, 1024]; [ ] T Indicates transpose. Represents the set of real numbers;

[0043] Then, after maintaining the dimension through a 128-node fully connected layer, the dimension is reduced to a 131072-dimensional vector through a flattening layer;

[0044] Finally, it is compressed into a 128-dimensional feature vector F through a 128-node fully connected layer. bi .

[0045] Furthermore, in step 9, the decision fusion process is as follows:

[0046] Step 9.1: Construct a fault type identification framework M={A1,A2,...,A...} based on Dempster-Shafer evidence theory. 13 The corresponding 13 types of failure modes;

[0047] Step 9.2, output the Softmax probability y mc (A x ) and y bi (A x ) as a source of evidence;

[0048] Calculate the basic trust level using the basic trust allocation formula, A. x Represents the fault mode type and is an element in the fault type identification framework M. The value range of x is [1, 13].

[0049] Step 9.3: Calculate the conflict factor K according to the Dempster-Shafer evidence theory; m1(B) and m2(C) are both sources of evidence in the evidence theory; among them, B and C are both focal elements with non-zero confidence.

[0050] Step 9.4: Calculate the confidence level of the basic trust assignment value before evidence synthesis;

[0051] The basic trust assignment function is defined. for:

[0052]

[0053] in: = |1 - max(Y mc | represents the network confidence error; the softmax output vector Y mc =[ , , . . ., ],m(A x ) indicates fault type A x The information source is n; n is the source number, and the value of n is [1, 13].

[0054] Calibration factor Where accuracy is the model validation accuracy, α is 0.85, and e is the natural constant;

[0055] Step 9.5: Fuse the evidence using the evidence synthesis formula and output the fault diagnosis result. m(A) represents the base probability assignment after fusion; where A is a focal element with non-zero confidence.

[0056] Basic probability assignment after fusion .

[0057] An intelligent diagnostic system based on the IMP self-test module, such as Figure 2 As shown, the intelligent diagnostic system is a voltage-current dual-mode diagnostic system, which includes a multi-source signal acquisition module, a parallel computing module, an evidence fusion module, and a diagnostic report generation module. The multi-source signals are imported into the parallel computing module for processing to obtain feature vectors. The feature vectors are used as evidence sources and input to the evidence fusion module for decision fusion through evidence theory. The diagnostic report generation module outputs the diagnostic results.

[0058] The multi-source signal acquisition module includes a precision sampling resistor array and a six-channel oscilloscope;

[0059] The parallel computing module is a dual-GPU acceleration unit; the dual-GPU acceleration unit includes a GPU-accelerated dilated convolution unit and a GPU-accelerated bidirectional long short-term unit.

[0060] The evidence fusion module is an embedded FPGA processor; it sequentially performs Dempster-Shafer fusion and calculates m(A) x Solve for the conflict quantity K.

[0061] The technical effects of this invention are as follows:

[0062] This invention employs a multi-source heterogeneous fusion method. It extracts and downsamples the voltage and current signals of the circuit under test (DUT) of the self-test module of an airborne integrated mission processor. The voltage data slices are then input into a dilated convolutional network, and the current data slices are input into a bidirectional long-short-time network. The feature vectors obtained from the two networks are fused using evidence theory to output the final diagnostic result. This invention achieves intelligent fault diagnosis of the self-test module of the airborne integrated mission processor, with an average diagnostic accuracy of 98.5%. It effectively realizes heterogeneous multi-source information fusion, significantly suppresses false alarms in the airborne integrated mission processor, and is beneficial for health management and predictive maintenance technology support for complex airborne systems. This invention features multi-source complementarity, with voltage-current heterogeneous sensing covering circuit impedance characteristics and energy transfer features, solving the problem of insufficient representation by a single signal source. It has a feature decoupling mechanism, with dilated convolution specifically targeting spatial anomalies and Bi-LSTM focusing on temporal evolution patterns to avoid feature interference. It also features innovative fusion decision-making: DS evidence theory effectively handles sensor conflicts, improving diagnostic confidence by 35%±8%. Attached Figure Description

[0063] Figure 1 This is a flowchart of the diagnostic framework of the present invention;

[0064] Figure 2 This is a schematic diagram of the intermittent fault intelligent diagnosis system of the present invention;

[0065] Figure 3 This is a confusion matrix diagram of the diagnostic results of the present invention. Detailed Implementation

[0066] Taking a quad op-amp dual second-order high-pass filter (cutoff frequency 35kHz) as an example, the diagnostic framework is as follows: Figure 1 As shown, the diagnostic system structure diagram is as follows: Figure 2 The implementation process of the solution is detailed below:

[0067] (I) In the hardware preparation phase, the AD8676 low-noise operational amplifier was selected to build the circuit. The voltage probe used was a Tektronix P5200A (50MHz bandwidth), and the sampling resistor was a Vishay WSLP2726 (0.1Ω±0.1%). Data acquisition was performed using an NI PXIe-5162 six-channel oscilloscope system with an anti-aliasing filter (cutoff frequency 375kHz). Fault injection was performed to simulate device failure states using a programmable relay matrix.

[0068] (II) Data Collection Standards

[0069] Normal state acquisition: Input a 10kHz~50kHz sweep frequency signal, record for 3 seconds at a 750kHz sampling rate.

[0070] Fault mode settings:

[0071] R0 failure: 0Ω short circuit, 1MΩ open circuit, intermittent contact (cycle 20ms±5ms)

[0072] U1 failure: Gain +10% drift, -30% drift, intermittent drift (±5%~±15%)

[0073] C1 failure: nominal value short circuit, open circuit, ±20%~±50% capacitance fluctuation.

[0074] Each acquisition includes 4 channels of data: U1 voltage, U4 voltage, C2 current, and U4 current.

[0075] (III) Model Building Process

[0076] Dilated Convolutional Network Parameters: Input Layer: 1024×1, Convolutional Layer 1: kernel=(3,1), filters=64, dilation=1, padding='valid', Expanded layer by layer to Convolutional Layer 9: dilation=256, Flattened Layer: Output 256 dimensions, Fully Connected Layer: 128 units, ReLU activation

[0077] Bidirectional Long Short-Term Memory (Bi-LSTM) network parameters: Input layer: 1024×1, Bi-LSTM network layer: units=128, return_sequences=True, Fully connected layer 1: 128 units, preserving the time dimension, Flattened layer: output 131072 dimensions, Fully connected layer 2: 128 units

[0078] Fusion classifier parameters: Feature concatenation layer: 256-dimensional input, Weighting layer: 26-unit fully connected, ReLU activation, Output layer: 13-unit Softmax

[0079] (iv) Training and verification process

[0080] 1. Dataset composition: Total number of samples: 8190 groups (630 groups / fault type), divided into training / test sets in a 7:3 ratio.

[0081] 2. Optimizer: Adam (lr=0.001), batch size 32, early stopping threshold 20 epochs

[0082] 3. Evaluation metrics: accuracy, recall, balanced score, and specificity.

[0083] (v) Diagnostic Examples

[0084] Table 1. CUT Diagnostic Results: Four Categories of Performance

[0085] <![CDATA[Failure Label sign > Fault codes accuracy <![CDATA[specific sex > Recall rate Balanced fraction F0 NF 1.000 1.000 1.000 1.000 F1 RIF1 0.973 0.997 0.989 0.981 F2 RIF2 0.989 0.999 0.978 0.983 F3 RIF3 0.984 0.998 0.974 0.989 F4 RPF 1.000 1.000 1.000 1.000 F5 UIF1 0.988 0.999 0.950 0.969 F6 UIF2 0.973 0.997 0.994 0.984 F7 UIF3 1.000 1.000 0.983 0.992 F8 UPF 1.000 1.000 1.000 1.000 F9 CIF1 0.956 0.996 0.956 0.956 F10 CIF2 0.972 0.998 0.961 0.966 F11 CIF3 0.973 0.998 1.000 0.986 F12 CPF 1.000 1.000 1.000 1.000

[0086] This patented method was validated using 2305 blind test samples (including normal operating conditions and three types of component failures), achieving an overall average diagnostic accuracy of 98.50%, significantly superior to existing similar solutions. Diagnostic results are as follows: Figure 3 As shown, the classification performance is as follows:

[0087] Normal state identification: All 180 normal samples were correctly classified, achieving a 100% recognition rate with no false alarms. This demonstrates the complete reliability of this solution in determining non-fault states.

[0088] Resistor R0 fault diagnosis: For intermittent and permanent faults of different severity in R1 components, the average recognition rate is 98.37%. Among them, the recognition rate of permanent PF faults is 100%, the recognition rate of severe intermittent faults of IF3 level is the highest (99.1%), and the recognition rate of mild faults of IF1 level is the lowest (97.2%). The main misjudgment is due to the characteristic overlap between IF1 and IF2 levels (false recognition rate 1.8%).

[0089] Operational amp U1 fault diagnosis: The identification rate for permanent faults (PF) in U1 reaches 100%; in intermittent fault diagnosis, the identification rate for minor faults at the UIF1 level is improved to 95.3% through heterogeneous information fusion. The key improvement lies in capturing the harmonic distortion characteristics caused by gain drift in the current signal, reducing the false identification rate of different fault levels (UIF1 / RIF1 / CIF1) to 2.1%-3.8%.

[0090] Capacitor C1 fault diagnosis: The average identification rate of the three types of intermittent faults (IF9-IF11) and permanent faults (PF12) of C1 component is 97.25%. Among them, the identification rate of PF12 is 100%, the identification rate of severe capacitance fluctuation of IF11 is 99.7%, and the identification rate of mild fault of IF9 is the lowest (96.1%). False judgments are concentrated on the cross-confusion of different severity of the same fault (e.g., the proportion of IF9 being misjudged as IF10 is 2.3%).

[0091] The timing control of the voltage-current dual-mode diagnostic system meets the following requirements: data acquisition time ≤ 3 seconds, feature extraction time ≤ 10ms, and evidence fusion time ≤ 5ms.

[0092] The multi-source information acquisition module is designed to overcome the bottleneck of single signal type perception. In the physical implementation of the airborne IMP self-test module circuit, the voltage monitoring point covers the output terminal U1 of the first-stage amplifier, the bandpass node BP, and the circuit output terminal U4.

[0093] The current detection point is set at the charging and discharging path of the compensation capacitor C2 and at the U4 node of the circuit output terminal, and a 0.1Ω±0.1% precision resistor is used for sampling; the sampling system is equipped with a six-channel synchronous acquisition card with a sampling rate of 750kHz and a synchronization accuracy error of ≤5ns, which meets the microsecond-level fault capture requirements;

[0094] Signal characteristic pattern: When U1 experiences early degradation, i.e., gain drift of +5%, the voltage signal V of U1... U1(t) Within a 3.58ms window, a dip fluctuation with an amplitude of 8% of the nominal value occurs, while the C2 current signal I_c2(t) simultaneously exhibits nonlinear distortion, i.e., the total harmonic distortion (THD) increases from 1.2% to 9.8%. When poor contact of resistor R0 causes intermittent open circuit, the voltage fluctuation amplitude of U4 changes by only 4%, but the U4 node current I... U4(t) A zero-value state was observed for 13.2 ms; this shows that the current signal has a more significant characteristic representation ability of load characteristic changes.

[0095] A parallel processing architecture is established to address the feature decoupling problem of heterogeneous signals. The voltage feature extraction path employs a nine-layer dilated convolutional network. The input 1024×1 voltage sequence is processed through a first 3×1 convolution with a dilation factor d1=1, yielding a 1024×64 feature map. The second convolution has a dilation factor d2=2, outputting a 1020×64 feature map. The layers are increased to the ninth layer with d=256, outputting a 4×64 feature matrix. Key design strategies include: eliminating pooling layers to avoid feature loss; exponentially expanding the receptive field to R9=1023; covering the complete signal cycle; and using ReLU activation and batch normalization between layers to accelerate convergence. The final feature vector F... mc Output from a 128-node fully connected layer.

[0096] The current feature extraction path is based on a bidirectional long short-term memory network; the input sequence is divided into 1024 time steps. In the bidirectional long short-term memory network, the forward unit calculates the hidden state, and the backward unit calculates the hidden state; the sequence is then concatenated to generate [h]. jf , h jb ] ∈ 256 h jf h jb The concatenated feature vectors, j=1,2,...,1024, are reduced to a 1024×128 feature matrix through a 128-node fully connected layer; after flattening, the output F is passed through another 128-node fully connected layer. bi The optimization of bidirectional long short-term memory network parameters includes: a time axis sliding window overlap rate of 50% and a dropout rate of 0.2 to prevent overfitting.

[0097] This solution achieves three breakthrough improvements through the fusion of voltage-current heterogeneous features: the sensitivity of intermittent fault level identification is increased by 37%, the recognition rate of U1 minor fault (UIF1) is increased from 92.1% to 95.3%, the cross-component misjudgment rate is reduced by 81%, the misjudgment rate of RIF1 and CIF1 is reduced from 4.9% to 0.9%, and the ability to distinguish the severity of faults is improved: the internal identification gradient of IF level faults reaches 3.2% (the difference in recognition rate between IF1 and IF3).

[0098] The effectiveness of this scheme is verified through comparative experiments, such as... Figure 3 As shown:

[0099] 1. Diagnostic accuracy: In 2340 blind test samples (including 180 normal samples), the overall accuracy was 98.50% ± 0.68% (standard deviation of ten repeated experiments). Specifically, for intermittent faults at stage IF1 of amplifier U1 (gain drift 5%~10%), the recognition rate increased from 89.3% with the traditional CNN method to 96.7%; the recognition rate for stage IF3 faults in capacitor C1 (capacitance fluctuation 30%~50%) reached 100%.

[0100] 2. In terms of false alarm suppression: the normal sample recognition rate is 100%, and the fault specificity index reaches 0.996, which is better than similar multi-source fusion schemes (0.982) and feature-level fusion schemes (0.975).

[0101] 3. In terms of computational efficiency: the number of parameters in dilated convolution is 27% of that in CNN, and the training time is reduced to one-third of the original; the dual-path parallel processing in the inference stage reduces the time for a single diagnosis to 15ms, meeting the 50ms real-time response requirement of avionics equipment.

Claims

1. A smart diagnostic method based on an IMP self-test module, characterized in that, The intelligent diagnostic method includes the following steps: Step 1, create a fault model classification; Intermittent faults of the airborne IMP self-test module are classified into circuit component resistor R faults, capacitor C faults, and operational amplifier U faults; fault states are classified into normal NF, intermittent fault IF, and permanent fault PF. The intermittent faults (IFs) are classified into three levels according to their severity: minor intermittent faults (IF1), moderate intermittent faults (IF2), and severe intermittent faults (IF3). Get the set of fault types The set of fault types Includes 13 types of failure modes; ； Among them, RIF1 indicates a minor intermittent resistor fault, RIF2 indicates a moderate intermittent resistor fault, RIF3 indicates a severe intermittent resistor fault, RPF indicates a permanent resistor fault, CIF1 indicates a minor intermittent capacitor fault, CIF2 indicates a moderate intermittent capacitor fault, CIF3 indicates a severe intermittent capacitor fault, CPF indicates a permanent capacitor fault, UIF1 indicates a minor intermittent op-amp fault, UIF2 indicates a moderate intermittent op-amp fault, UIF3 indicates a severe intermittent op-amp fault, and UPF indicates a permanent op-amp fault. Step 2: Deploy voltage monitoring points; The voltage monitoring point includes the input terminal U of the IMP self-test module circuit. in The circuit consists of the first-stage amplifier output terminal U1, the bandpass node BP, the key node of the RC network C1, the key node of the RC network R0, and the circuit output terminal U4. Step 3: The multi-source signal acquisition module uses a voltage monitoring sensor to acquire the voltage signal V(t) and the voltage drop Vr(t) across the sampling resistor; Step 4: Calculate the current signal I(t); In the charging and discharging path C of energy storage capacitor C1 path With load terminal L path power load R P A precision sampling resistor is connected in series, and a compensation capacitor C2 is set simultaneously, with C2 connected in parallel across the sampling resistor; the current signal I(t) is obtained according to the current conversion formula I(t)=Vr(t) / Rr, where Rr=0.1Ω±0.1%; Step 5: Use an anti-aliasing filter to acquire the voltage signal V(t) and generate a time series dataset; Step 6: The voltage signal V(t), the resistance voltage drop Vr(t), and the current signal I(t) are downsampled as raw signal data to obtain the corresponding slice signals. Step 7: Input the voltage signal V(t) slice signal into the parallel computing module, and use a nine-layer dilated convolutional network to extract spatial features, obtaining a 128-dimensional feature vector F. mc ; Step 8: Similarly, the sliced ​​current signal I(t) is input into the parallel computing module and bidirectional long short-term memory network to obtain the 128-dimensional feature vector F. bi ; Step 9: The evidence fusion module uses evidence theory to analyze the 128-dimensional feature vector F. mc and 128-dimensional eigenvector F bi After decision fusion, the results are input into the diagnostic report generation module, which then outputs the fault diagnosis result A*. The decision fusion process is as follows: Step 9.1: Construct a fault type identification framework M={A1,A2,...,A...} based on Dempster-Shafer evidence theory. 13 The corresponding 13 types of failure modes; Step 9.2, output the Softmax probability y mc (A x ) and y bi (A x ) as a source of evidence; The basic trust level is calculated using the basic trust assignment function, A. x Represents the fault mode type and is an element in the fault type identification framework M. The value range of x is [1, 13]. Step 9.3: Calculate the conflict factor K according to the Dempster-Shafer evidence theory; Both m1(B) and m2(C) are sources of evidence in the evidence theory; among them, B and C are focal elements with non-zero confidence. Step 9.4: Calculate the confidence level of the basic trust assignment value before evidence synthesis; The basic trust allocation function for: （3）； in: = |1 - max(Y mc | represents the network confidence error; the softmax output vector Y mc =[ , , . . ., ],m(A x ) indicates fault type A x The information source is n; n is the source number, and the value of n is [1, 13]. Calibration factor Where accuracy is the model validation accuracy, α is 0.85, and e is the natural constant; Step 9.5: Fuse the evidence using the evidence synthesis formula and output the fault diagnosis result. m(A) represents the base probability assignment after fusion; where A is a focal element with non-zero confidence. Basic probability assignment after fusion .

2. The intelligent diagnostic method based on the IMP self-test module according to claim 1, characterized in that, In step 3, the voltage signal V(t) and the voltage drop across the sampling resistor Vr(t) are synchronously acquired at a sampling rate of 750kHz.

3. The intelligent diagnostic method based on the IMP self-test module according to claim 1, characterized in that, In step 4, the sampling resistor has an accuracy class of 0.

02.

4. The intelligent diagnostic method based on the IMP self-test module according to claim 1, characterized in that, In step 5, the passband cutoff frequency of the anti-aliasing filter is 375kHz, and the continuous acquisition time is 3 seconds.

5. The intelligent diagnostic method based on the IMP self-test module according to claim 1, characterized in that, In step 6, the downsampling settings are as follows: The sampling rate was first reduced from 750kHz to 250kHz before data acquisition, resulting in 600 sets of samples. During data acquisition, the acquisition lasted for 3 seconds, with a signal period of approximately 4ms; each sample group had a length of 1024 points.

6. The intelligent diagnostic method based on the IMP self-test module according to claim 1, characterized in that, In step 7, the layer parameters of the nine-layer dilated convolutional network include the receptive field and the dilation factor; Employing a zero-fill strategy to maintain output size stability, receptive field The extended recurrence formula is: （1）； （2）； Where l represents the number of dilated convolution layers; Indicates the output of the current layer; represents the input of the current layer; k represents the size of the convolution kernel; the receptive field of the first layer is R1=3; the ninth layer is 1023 points; the size of the output feature matrix of the nine-layer dilated convolutional network is decreased by formulas (1) and (2).

7. The intelligent diagnostic method based on the IMP self-test module according to claim 1, characterized in that, In step 7, the 128-dimensional feature vector F mc The acquisition process is as follows: First, the voltage signal V(t) slice signal is passed through a nine-layer dilated convolutional network to output a 4×64 feature matrix; Each layer of the dilated convolutional network is configured with 64 3×1 convolutional kernels with a fixed stride of 1 and an inflation factor d. l =2^(l-1) doubling layer by layer, where l represents the number of dilated convolutional layers; Then, the 4×64 feature matrix is ​​flattened and compressed into a 128-dimensional feature vector F by a 128-node fully connected layer. mc .

8. The intelligent diagnostic method based on the IMP self-test module according to claim 1, characterized in that, In step 8, the 128-dimensional feature vector F bi The acquisition process is as follows: First, the 1024×1 current signal I(t) slice signal is processed by a 128-processing-unit bidirectional long short-term memory network to synchronously calculate the forward propagation state and the backward propagation state, outputting the hidden state matrix, which is then reduced to a 1024×128 spatiotemporal feature matrix H through a 128-node fully connected layer. （3）； Among them, h j Let represent the feature vector at time step j, where j ranges from [1, 1024]; [ ] T Indicates transpose. Represents the set of real numbers; Then, after maintaining the dimension through a 128-node fully connected layer, the dimension is reduced to a 131072-dimensional vector through a flattening layer; Finally, it is compressed into a 128-dimensional feature vector F through a 128-node fully connected layer. bi .

9. A diagnostic system for performing the intelligent diagnostic method according to claims 1 to 8, wherein the intelligent diagnostic system is a voltage-current dual-mode diagnostic system, comprising, in sequence, a multi-source signal acquisition module, a parallel computing module, an evidence fusion module, and a diagnostic report generation module; The multi-source signal acquisition module acquires multi-source signals, including voltage signals and current signals. Multi-source signals are imported into a parallel computing module for processing to obtain feature vectors; features Vectors are input as evidence sources to the evidence fusion module, where decision fusion is performed using evidence theory; the diagnostic report generation module outputs the diagnostic results. The multi-source signal acquisition module includes a precision sampling resistor array and a six-channel oscilloscope; The parallel computing module is a dual-GPU acceleration unit; the dual-GPU acceleration unit includes a GPU-accelerated dilated convolution unit and a GPU-accelerated bidirectional long short-term unit. The evidence fusion module is an embedded FPGA processor; it sequentially performs Dempster-Shafer fusion and calculates m(A) x Solve for the conflict quantity K.

Images