TCSNN-based aero-engine gas path component recession prediction system

Abstract

The invention discloses a TCSNN-based aero-engine gas path component recession prediction system, and relates to the field of artificial intelligence. The TCSNN-based aero-engine gas path component recession prediction system comprises a data input module, a performance prediction module and a data output module which are deployed on an FPGA (Field Programmable Gate Array), the exhaust temperature reflecting the performance of the whole machine is predicted based on the gas path performance parameters of the multiple components; according to the system, the TCSNN network model is adopted, the TCN is improved, and the anti-interference capability and the prediction accuracy of the network model are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and particularly to a degradation prediction system for aero-engine gas path components based on TCSNN. Background Art

[0002] The performance prediction and trend analysis of aero-engines are to predict the future change trends of engines and potential faults of key components by analyzing current and historical data. The aero-engine system is extremely complex, and its fault types mainly include gas path faults, mechanical faults, control system faults, vibration faults, etc. Among them, the gas path system is responsible for regulating and controlling the airflows of key components such as compressors, combustors, and turbines, and its stability directly affects the working efficiency and safety of the engine. Under cruise conditions, the main gas path performance parameters of the engine include thrust, fuel consumption, exhaust gas temperature, etc. The change trends of key indicators can reflect the performance degradation of the engine. However, since these performance parameters are difficult to obtain in real time or there are measurement difficulties, therefore, predicting the health status and working efficiency of the engine through the performance parameters of gas path components that can be directly measured has become the key to research. Through accurate trend prediction, signals of performance degradation can be detected in a timely manner, and then a more scientific maintenance and repair strategy can be realized.

[0003] With the development of aviation technology, traditional data analysis methods based on experience and maintenance history can no longer meet the complexity and high-precision requirements of modern aero-engines. Data-driven models and algorithms predict the performance of aero-engines in real time through neural network models, avoiding the complexity of traditional physical modeling. In view of the characteristics of the time-series data collected by sensors, it is required that the neural network model used has efficient time-series modeling capabilities. Among various data-driven methods, the Temporal Convolutional Network (TCN) has been widely used in gas path performance prediction due to its strong time-series modeling capabilities. Through causal convolution and dilated convolution structures, TCN can effectively capture long-span dependencies. However, there are still some problems in the actual application of TCN. TCN has a high demand for computing resources. Especially when dealing with large-scale data, the computing efficiency may become a bottleneck. When dealing with non-linear dynamic changes, there are also problems that complex time-series features cannot be fully captured, resulting in limited prediction accuracy. At the same time, traditional TCN models deployed on the CPU (Central Processing Unit) / GPU (Graphics Processing Unit) platform cannot meet the real-time requirements of engine performance prediction.

[0004] By introducing spiking neurons into the TCN, the computational tasks can be sparsified, reducing resource occupancy while maintaining a high sensitivity to the dynamic changes of time-series signals. The TCN combined with spiking neurons is more suitable for hardware acceleration through FPGA (Field Programmable Gate Array) in terms of resource occupancy and power consumption. Compared with the neural network models deployed on the CPU / GPU platform, it can also obtain a higher computational parallelism to achieve real-time prediction of engine performance. Therefore, it is of practical significance to explore an aero-engine gas path component degradation prediction system based on the Temporal Convolutional Spiking Neural Network (TCSNN). Summary of the Invention

[0005] The present invention proposes an aero-engine gas path component degradation prediction system based on TCSNN. This system is based on the time-series signals of the aero-engine gas path performance parameters to predict the overall performance of the engine. The present invention improves the temporal convolutional network, uses causal convolution and residual connection to capture the long-term dependencies in the data, and uses spiking neurons to encode the temporal information, thereby further improving the time resolution and learning ability and enhancing the accuracy of the network model prediction. The aero-engine gas path component degradation prediction system based on TCSNN is deployed on the FPGA to meet the real-time requirement of aero-engine detection through hardware acceleration technology.

[0006] Technical solution of the present invention:

[0007] An aero-engine gas path component degradation prediction system based on TCSNN, the aero-engine gas path component degradation prediction system based on TCSNN includes a data input module, a performance prediction module and a data output module deployed on the FPGA;

[0008] The data input module includes a sensor module, a UART serial port receiving module, a data processing module, and a line buffer RAM module; the composition and specific functions of each module are as follows:

[0009] Sensor module: includes various types of sensors for collecting the original data of the parameters of each gas path component changing with time;

[0010] UART serial port receiving module: used to receive the original data of the parameters of each gas path component changing with time collected by the sensor module, and generate feature data after parsing according to the UART protocol;

[0011] Data processing module: It is used to process the feature data parsed according to the UART protocol by the UART serial port receiving module to obtain a feature sequence; the data processing module is deployed on the PS side of the FPGA, and the obtained feature sequence is transmitted from the PS (Processing System) side to the PL (Programmable Logic) side through the AXI bus;

[0012] Row buffer RAM module: It is used to buffer the feature sequence processed by the data processing module;

[0013] The performance prediction module includes a weight loading module, a weight ROM module, a temporal convolutional network calculation module, a data flow control module, a spiking neuron calculation module, and a fully connected calculation module; the composition and specific functions of each module are as follows:

[0014] Weight loading module: The temporal convolutional network calculation module sends a weight address index to the weight loading module. The weight address index includes the count signal of the number of convolutional cycles within a single layer (Count of Convolution Cycles within Per-Layer, C3PL) and the count signal of the overall convolutional layer cycles (Count of Convolutional Layer Cycles, CCLC). The weight loading module generates the address of the weight parameter to be read according to the weight address index;

[0015] Weight ROM module: It is used to write the weight parameters obtained from TCSNN training into the BRAM (Block RAM) of the FPGA in the form of a COE file, and output the required weight parameters to the temporal convolutional network calculation module according to the address of the weight parameter generated by the weight loading module;

[0016] Temporal convolutional network calculation module: It includes an input data cache RAM unit (Input Data RAM, IDRAM), a convolution result cache RAM unit (Convolution Results RAM, CRRAM), and a convolution and residual general calculation (Convolution-Residual General Calculation, C-RGC) unit;

[0017] Input data cache RAM unit: It is used to cache the feature sequence from the row buffer RAM module in the data input module;

[0018] Convolution result cache RAM unit: It is used to cache the intermediate result data after the convolution and residual general calculation unit performs convolution calculation and residual calculation;

[0019] Convolution and residual general computing unit: According to the shape of TCSNN and the DSP (Digital Signal Processor) and LUT (Look-Up Table) resources of FPGA, the convolution calculation parallelism is designed. The convolution and residual general computing unit reads the feature sequence in the input data cache RAM unit and the intermediate result data in the convolution calculation result cache RAM unit according to the designed convolution calculation parallelism, and reads the weight parameters in the weight ROM module that matches the feature sequence to perform convolution and residual calculations. During the convolution and residual calculation process, the convolution calculation process of each layer needs to continuously update the convolution cycle count signal within the single layer to refresh the weight parameters. After the convolution calculation of each layer is completed, the overall cycle count signal of the convolution layer is updated, and the intermediate result number of the convolution calculation of the current layer is updated. The data is cached in the convolution calculation result cache RAM unit. When the convolution calculation of the next layer starts, the intermediate result data of the convolution calculation of the previous layer is directly read from the convolution calculation result cache RAM unit to perform the convolution calculation of the next layer. When the convolution calculation of all layers is completed, the convolution and residual general calculation unit performs residual calculation: read the feature sequence from the input data cache RAM unit, perform convolution calculation on the feature sequence and the refreshed weight parameter, and perform weighted addition on the result of the convolution calculation and the convolution calculation result of the last layer that has been cached in the convolution calculation result cache RAM unit. At this point, all calculations of the entire time convolution network calculation module are completed.

[0020] Pulse neuron calculation module: The pulse neuron calculation module includes Z pulse neurons, according to the time step parameters designed for each pulse neuron Build The pipeline structure is composed of a set of LIF (LeakyIntegrate and Fired) model operators. The membrane potential of the initial cache in the first-level pipeline structure is set to zero, and the membrane potential is transmitted to the next-level pipeline structure or an output pulse is generated according to the result signal output by the time convolution network calculation module; the pulse neuron calculation module converts the result signal output by the time convolution network calculation module into a pulse signal based on the firing sequence of the pulse neuron; the LIF model operator is explained by the following formula:

[0021] (1)

[0022] (2)

[0023] Where: represents the time constant of the membrane, represents the membrane potential of the spiking neuron, represents the resting potential of a spiking neuron, represents the membrane resistance, Indicates the change of the externally input current over time ; Equation (1) describes the effect of the leakage of the resting potential and the externally input current on the membrane potential; Indicates the moment when the spiking neuron fires a spike, Indicates the time when the membrane potential of the spiking neuron approaches 0 after firing a spike The membrane potential of the spiking neuron within, Indicates the reset value of the membrane potential of the spiking neuron after firing a spike; Equation (2) describes that after the spiking neuron fires a spike at the moment , it will reset its membrane potential to within the time approaching 0 ;

[0024] Data flow control module: The pulse signal output by the spiking neuron calculation module needs to be input into the time convolutional network calculation module C times for convolution calculation and residual calculation, and the spiking neuron calculation module converts it into a pulse signal. After the last time the time convolutional network calculation module finishes the calculation, it is processed and converted into a pulse signal by the spiking neuron calculation module and then sent to the fully connected calculation module for calculation; The data flow control module controls the output flow direction between each module in the performance prediction module according to the convolution loop count signal within a single layer and the overall convolution layer loop count signal;

[0025] Fully connected calculation module: The fully connected calculation module is responsible for mapping the pulse signal finally obtained by the spiking neuron calculation module to the final output space, and mapping the pulse signal to the predicted value of the exhaust gas temperature through a two-layer fully connected network.

[0026] The described data output module includes a UART serial port sending module and a threshold judgment module;

[0027] UART serial port sending module: Used to pack the predicted value of the exhaust gas temperature obtained by the performance prediction module according to the UART protocol and send it to the host computer through the serial port, for displaying the predicted value of the exhaust gas temperature and safety alarm;

[0028] Threshold judgment module: Used to monitor whether the predicted value of the exhaust gas temperature is higher than the safety threshold. If the predicted value of the exhaust gas temperature is higher than the safety threshold, a safety alarm is given.

[0029] Before deploying the aero-engine gas path component degradation prediction system based on TCSNN on the FPGA, it is necessary to train TCSNN. The process is as follows:

[0030] Step 1: Collect the original data of the parameters of each gas path component of the aero-engine changing with time while maintaining the cruise mode through sensors; perform preprocessing on the original data, including normalization, noise reduction, and data augmentation, to obtain the preprocessed gas path performance parameters; perform correlation analysis on each gas path performance parameter with the exhaust gas temperature, and use the gas path performance parameters with the correlation function value greater than the set threshold as the feature sequence input to the TCSNN; divide the feature sequence into a training set, a validation set, and a test set according to a certain proportion.

[0031] Further, step 1 specifically includes the following steps:

[0032] Step 1.1: Collect the original data of the low-pressure rotor speed, high-pressure rotor speed, high-pressure compressor outlet pressure, oil pressure, compressor outlet temperature, and guide angle changing with time while the aero-engine maintains the cruise mode through sensors at a fixed sampling frequency.

[0033] Step 1.2: Use normalization processing to scale the original data to a data sequence between ; use the wavelet packet threshold denoising method to perform denoising processing on the normalized data sequence to obtain a denoised data sequence; use the sliding overlapping sampling method to perform data augmentation on the denoised data sequence.

[0034] Even further, step 1.2 specifically includes the following steps:

[0035] Step 1.2.1: Perform normalization processing on the original data to scale the original data to between to obtain a data sequence with the number of samples being ; ;

[0036] Each item in the data sequence is calculated as follows:

[0037]

[0038] where is the original data, is the sequence of the original data, is the maximum value in the sequence of the original data, is the minimum value in the sequence of the original data;

[0039] Step 1.2.2: Use wavelet basis to perform layers of wavelet decomposition on the normalized data sequence to obtain layers of wavelet coefficients, where the wavelet coefficients obtained from each layer of wavelet decomposition include approximation coefficients (low-frequency components) and detail coefficients (high-frequency components); for the first layer to the The detail coefficients of the layer are screened using a soft threshold function; for the approximate coefficients of the layer and the detail coefficients from the first layer to the layer after being screened by the soft threshold function, perform inverse wavelet transform to reconstruct the denoised data sequence ; where is the denoised data;

[0040] Step 1.2.3: Adopt the sliding overlapping sampling method and use a fixed sample length and overlapping length to divide the denoised data sequence into sample sequences, denoted as:

[0041]

[0042] where is the denoised data of the th item in the denoised data sequence ;

[0043] Step 1.3: Calculate the mutual information function values between the performance parameters of each gas path and the exhaust temperature after being processed in Step 1.2. When the mutual information function value is greater than the set threshold, it is used as the feature sequence input to TCSNN. When the mutual information function value is less than the set threshold, it is not used as the feature sequence of TCSNN;

[0044] The mutual information function between the performance parameters of each gas path and the exhaust temperature

[0045]

[0046] In the formula, is the denoised data sequence, is the exhaust temperature sequence , , and are the entropies of , and respectively, and there is:

[0047]

[0048]

[0049]

[0050] In the formula, represents 's probability density, represents the probability density of is and the joint probability density of

[0051] Step 1.4: Divide the feature sequence obtained in Step 1.3 into a training set, a validation set, and a test set according to a ratio;

[0052] Step 2: Construction and training of the temporal convolutional spiking neural network TCSNN;

[0053] TCSNN is obtained by adding a spiking neuron calculation module to the temporal convolutional network TCN; the temporal convolutional network TCN contains C layers of residual network structures, and each layer of residual network structure consists of two convolutional layers and one residual layer. The C layers of residual network structures are used for feature extraction and temporal modeling of the input feature sequence; the spiking neuron calculation module includes X spiking neurons, and each spiking neuron is placed between two adjacent layers of residual network structures to play a connecting role. At the same time, based on the firing timing of the spiking neurons, the continuous signal output by the residual network structure is converted into a spiking signal input and transmitted to the spiking neurons to further optimize the prediction result;

[0054] The training process of TCSNN is as follows: First, initialize the hyperparameters and weights of TCSNN; then input the training set into TCSNN in batches for training until TCSNN converges; by adjusting the hyperparameters, obtain the optimal TCSNN; specifically as follows:

[0055] Step 2.1: Initialize the hyperparameters and weights of TCSNN;

[0056] Step 2.2: Input the training set into TCSNN in batches and output the predicted value of the exhaust temperature of the aeroengine. Compare the predicted value of the exhaust temperature with the actual value, calculate the loss value Loss through the mean squared error function (MSE), and calculate the loss function for the weights of the convolutional layer and the residual layer gradients, and use the gradient descent algorithm to backpropagate and update the weights of each convolutional layer and residual layer of TCSNN;

[0057] Furthermore, the specific formula for the loss value Loss is:

[0058]

[0059] where is the number of samples, is the actual value, is the predicted value of TCSNN;

[0060] Furthermore, the loss function with respect to the weights has the following gradient formula:

[0061]

[0062] Step 2.3: Determine whether the loss value is less than the threshold . When the loss value is less than the threshold , it is considered that TCSNN reaches the optimum and the training stops. Otherwise, adjust the hyperparameters and repeat Step 2.2;

[0063] Step 3: Evaluate the storage and computing resource requirements of TCSNN on FPGA; load the weight parameters obtained by training TCSNN into FPGA in the format of a COE file for storage;

[0064] Step 3.1: According to the data format of the weight parameters used in the deployment in FPGA, evaluate the usage amounts of storage resources and logical computing resources of TCSNN and the deployment. Among them, use the on-chip DSP resources to calculate the tensor multiplication in convolution and matrix calculations, and use LUT to build the basic unit of TCSNN;

[0065] Step 3.2: Write the weight parameters obtained by training TCSNN into a COE file for the weight ROM module on FPGA to read for the forward propagation of TCSNN.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] The present invention proposes aeroengine gas path component degradation prediction based on TCSNN, and predicts the exhaust gas temperature reflecting the overall engine performance based on the gas path performance parameters of multiple components; the system adopts the TCSNN network model, improves the TCN, and enhances the anti-interference ability and prediction accuracy of the network model;

[0068] In the data preprocessing stage, double optimizations of wavelet packet threshold denoising and mutual information feature screening are introduced. The original signal is decomposed at multiple scales using the 5th-order Daubechies wavelet basis, and high-frequency noise is accurately removed through the soft threshold function to improve the signal-to-noise ratio. The mutual information method is used to screen the feature parameters strongly correlated with the exhaust gas temperature (threshold > 0.9), eliminate redundant interference, and improve the prediction stability of the model under complex working conditions;

[0069] In the network model used in the present invention, causal convolution is employed. Through the design of one-way convolution kernels, the causality of time-series signal processing is ensured, avoiding the problem of information leakage. Residual connections are used. Through cross-layer feature fusion, the vanishing gradient problem in deep networks is solved, thereby efficiently extracting time-series features and capturing long-term dependencies in the data. Spiking neurons model the dynamic changes of time-series signals. Utilizing their biologically inspired dynamic threshold mechanism, spike signals are triggered only in high-information regions, significantly reducing the redundant computational amount, further improving the time resolution and learning ability, and at the same time reducing the computational complexity through sparse activation;

[0070] The system is deployed on an FPGA to achieve low-latency computing. When performing convolution calculations, by designing a circular RAM storage structure, all intermediate calculation results share the same part of the on-chip RAM resources, which can greatly reuse and save on-chip RAM resources and reduce the resource overhead caused by frequent data transfer. The spiking neuron calculation module adopts a cascaded pipeline design. By parallelly processing the membrane potential updates of multiple time steps, the timing blocking problem in traditional serial calculations is avoided;

[0071] Combined with the parallel computing ability of the FPGA, the system can complete the synchronous processing of multi-channel data within a single clock cycle, and the overall prediction latency is less than 5 ms, meeting the real-time monitoring requirements of aeroengines. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a framework diagram of the aeroengine gas path component degradation prediction system based on TCSNN of the present invention;

[0073] Figure 2 is a schematic diagram of the network model structure of TCSNN of the present invention;

[0074] Figure 3 is a deployment diagram of the network model of TCSNN of the present invention on an FPGA development board. Among them, (a) is a deployment schematic diagram of the performance prediction module, and (b) is a deployment schematic diagram of the temporal convolutional network module;

[0075] Figure 4 is a schematic diagram of the design of the circular RAM storage structure of the present invention;

[0076] Figure 5 is a comparison diagram of the true value and the model value of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0077] The following further describes the specific implementation manners of the present invention in combination with the drawings and technical solutions.

[0078] The present invention proposes an aeroengine gas path component degradation prediction system based on TCSNN, as Figure 1As shown in the figure, the aero-engine gas path component degradation prediction system based on TCSNN includes a data input module, a performance prediction module, and a data output module deployed on the FPGA;

[0079] The data input module includes a sensor module, a UART serial port receiving module, a data processing module, and a line buffer RAM module; the composition and specific functions of each module are as follows:

[0080] Sensor module: It includes various types of sensors for collecting the original data of the parameters of each gas path component changing with time;

[0081] UART serial port receiving module: It is used to receive the original data of the parameters of each gas path component changing with time collected by the sensor module, and generate feature data after parsing according to the UART protocol;

[0082] Data processing module: It is used to process the feature data generated by the UART serial port receiving module according to the UART protocol parsing to obtain a feature sequence; the data processing module is deployed on the PS side of the FPGA, and the obtained feature sequence is transmitted from the PS (Processing System) side to the PL (Programmable Logic) side through the AXI bus;

[0083] Line buffer RAM module: It is used to cache the feature sequence processed by the data processing module;

[0084] The performance prediction module: As shown in (a) of Figure 3 it includes a weight loading module, a weight ROM module, a temporal convolutional network calculation module, a data flow control module, a spiking neuron calculation module, and a fully connected calculation module; the composition and specific functions of each module are as follows:

[0085] Weight loading module: The temporal convolutional network calculation module sends a weight address index to the weight loading module. The weight address index includes a per-layer convolutional cycle count signal (Count of Convolution Cycles within Per-Layer, C3PL) and a convolutional layer overall cycle count signal (Count of Convolutional Layer Cycles, CCLC). The weight loading module generates the address of the weight parameter to be read according to the weight address index;

[0086] Weight ROM module: It is used to write the weight parameters obtained by TCSNN training into the BRAM (Block RAM) of the FPGA in the form of a COE file, and output the required weight parameters to the temporal convolutional network calculation module according to the address of the weight parameter generated by the weight loading module;

[0087] Temporal Convolution Network Computing Module: As shown in (b) of Figure 3 , it includes an Input Data Buffer RAM Unit (InputData RAM, IDRAM), a Convolution Results Buffer RAM Unit (Convolution Results RAM, CRRAM), and a Convolution-Residual General Calculation (C-RGC) Unit;

[0088] Input Data Buffer RAM Unit: Used to buffer the feature sequence from the row buffer RAM module in the data input module;

[0089] Convolution Results Buffer RAM Unit: Used to buffer the intermediate result data after the convolution and residual calculations by the convolution and residual general calculation unit;

[0090] Convolution-Residual General Calculation Unit: Design the convolution calculation parallelism according to the shape of the TCSNN and the DSP (Digital Signal Processor) and LUT (Look-Up Table) resources of the FPGA. The convolution-Residual General Calculation Unit reads the feature sequence in the input data buffer RAM unit and the intermediate result data in the convolution results buffer RAM unit according to the designed convolution calculation parallelism, and at the same time reads the weight parameters in the weight ROM module that matches the feature sequence to perform convolution calculation and residual calculation; during the convolution calculation and residual calculation process, the convolution loop count signal within a single layer needs to be continuously updated to refresh the weight parameters during the convolution calculation of each layer; after the convolution calculation of each layer is completed, update the overall convolution layer loop count signal, and cache the intermediate result data of the current layer's convolution calculation in the convolution results buffer RAM unit. When the next layer's convolution calculation starts, directly read the intermediate result data of the previous layer's convolution calculation from the convolution results buffer RAM unit to perform the next layer's convolution calculation; when all layers of convolution calculations are completed, the convolution-Residual General Calculation Unit performs the residual calculation: reads the feature sequence from the input data buffer RAM unit, performs convolution calculation on the feature sequence and the refreshed weight parameters, and adds the result of the convolution calculation to the weighted sum of the convolution calculation result of the last layer that has been cached in the convolution results buffer RAM unit. Thus, all calculations of the entire temporal convolution network computing module are completed;

[0091] Spiking Neuron Computing Module: The spiking neuron computing module includes Z spiking neurons, and constructs according to the time step parameters designed for each spiking neuron build The pipeline structure of each level is constructed by a group of LIF (Leaky Integrate and Fired) model operators. The initial membrane potential cached in the first-level pipeline structure is set to zero. According to the result signal output by the time convolution network calculation module, the membrane potential is passed to the next-level pipeline structure or an output pulse is generated; the spiking neuron calculation module converts the result signal output by the time convolution network calculation module into a spiking signal based on the firing timing of the spiking neuron; the LIF model operator is explained by the following formula:

[0092] (1)

[0093] (2)

[0094] In the formula: represents the time constant of the membrane, represents the membrane potential of the spiking neuron, represents the resting potential of the spiking neuron, represents the membrane resistance, represents the change of the externally input current over time ; Formula (1) describes the effect of the leakage of the resting potential and the influence of the externally input current on the membrane potential; represents the moment when the spiking neuron fires a pulse, represents the time when the membrane potential of the spiking neuron approaches 0 after firing a pulse within the spiking neuron, represents the reset value of the membrane potential after the spiking neuron fires a pulse; Formula (2) describes that after the spiking neuron fires a pulse at the moment , its membrane potential will be reset to within the time that approaches 0;

[0095] Data flow control module: The spiking signal output by the spiking neuron calculation module needs to be input into the time convolution network calculation module C times for convolution calculation and residual calculation, and the spiking neuron calculation module converts it into a spiking signal. After the last time the time convolution network calculation module finishes the calculation, it is processed by the spiking neuron calculation module and converted into a spiking signal, and then sent to the fully connected calculation module for calculation; the data flow control module controls the output flow direction between the various modules in the performance prediction module according to the convolution loop count signal within a single layer and the overall convolution layer loop count signal;

[0096] Fully connected calculation module: The fully connected calculation module is responsible for mapping the spiking signal finally obtained by the spiking neuron calculation module to the final output space, and maps the spiking signal to the predicted value of the exhaust gas temperature through a two-layer fully connected network.

[0097] The described data output module includes a UART serial port sending module and a threshold judgment module;

[0098] UART serial port sending module: It is used to pack the predicted value of the exhaust temperature obtained by the performance prediction module according to the UART protocol and send it to the host computer through the serial port, for displaying the predicted value of the exhaust temperature and safety alarm;

[0099] Threshold judgment module: It is used to monitor whether the predicted value of the exhaust temperature is higher than the safety threshold. If the predicted value of the exhaust temperature is higher than the safety threshold, a safety alarm is given.

[0100] Before deploying the aero-engine gas path component degradation prediction system based on TCSNN on the FPGA, it is necessary to train the TCSNN. The process is as follows:

[0101] Step 1: Collect the original data of the parameters of each gas path component of the aero-engine changing with time while maintaining the cruise mode through sensors; preprocess the original data, and the preprocessing includes normalization, noise reduction and data augmentation processing to obtain the preprocessed gas path performance parameters; perform correlation analysis on each gas path performance parameter with the exhaust temperature, and use the gas path performance parameters with the correlation function value greater than the set threshold as the feature sequence input to TCSNN; divide the feature sequence into a training set, a validation set and a test set according to a certain proportion;

[0102] Furthermore, the specific steps of Step 1 are as follows:

[0103] Step 1.1: Set the sampling frequency to 32768Hz, and use sensors to collect the original data of the low-pressure rotor speed, high-pressure rotor speed, high-pressure compressor outlet pressure, oil pressure, compressor outlet temperature, and guide angle changing with time while the aero-engine maintains the cruise mode;

[0104] Step 1.2: Use normalization processing to scale the original data to a data sequence between; use the wavelet packet threshold denoising method to perform denoising processing on the normalized data sequence to obtain a denoised data sequence; use the sliding overlapping sampling method to perform data augmentation on the denoised data sequence, use the amount of data collected per second as the window length of overlapping sampling, and set the step size to 8192;

[0105] Even further, the specific steps of Step 1.2 are as follows:

[0106] Step 1.2.1: Perform normalization processing on the original data to scale the original data to between, and obtain a data sequence with the number of samples being ; ;

[0107] Data sequence Each item in has the following calculation formula:

[0108]

[0109] Among them, is the original data, is the sequence of the original data, is the maximum value in the sequence of the original data, is the minimum value in the sequence of the original data;

[0110] Step 1.2.2: Use the wavelet basis to perform -layer wavelet decomposition on the normalized data sequence, obtaining -layer wavelet coefficients, where the wavelet coefficients obtained by each layer of wavelet decomposition include approximation coefficients (low-frequency components) and detail coefficients (high-frequency components); use the soft threshold function to screen the detail coefficients from the first layer to the -th layer; perform wavelet inverse transform on the approximation coefficients of the -th layer and the detail coefficients from the first layer to the -th layer after being screened by the soft threshold function to reconstruct the denoised data sequence ; among them, is the denoised data;

[0111] Step 1.2.3: Adopt the sliding overlapping sampling method and adopt the fixed sample length and the overlapping length to divide the denoised data sequence into sample sequences, denoted as:

[0112]

[0113] Among them, is the denoised data of the -th item in the denoised data sequence ;

[0114] Step 1.3: Calculate the mutual information function values between each gas path performance parameter and the exhaust temperature after being processed by Step 1.2. When the mutual information function value is greater than the set threshold, it is used as the feature sequence input to TCSNN. When the mutual information function value is less than the set threshold, it is not used as the feature sequence of TCSNN;

[0115] The mutual information function between each gas path performance parameter and the exhaust temperature

[0116]

[0117] In the formula, is the noise-reduced data sequence, is the exhaust gas temperature sequence , , and are respectively , and 's entropy, and there is:

[0118]

[0119]

[0120]

[0121] In the formula, represents 's probability density, represents 's probability density, is and 's joint probability density;

[0122] Table 1 shows the mutual information function values of the exhaust gas temperature EGT and the low-pressure rotor speed , the high-pressure rotor speed , the pressure at the outlet of the high-pressure compressor , the oil pressure , the temperature at the outlet of the compressor , the guide angle :

[0123] Table 1 Mutual information function values

[0124]

[0125] Taking 0.9 as the correlation analysis threshold, , , 's mutual information function values are greater than the correlation analysis threshold and are used as effective parameters for predicting the exhaust gas temperature;

[0126] Step 1.4: Divide the feature sequence obtained in Step 1.3 into a training set, a validation set, and a test set according to the ratio of 7:1:2;

[0127] Step 2: Construction and training of the time series convolutional spiking neural network TCSNN;

[0128] As Figure 2As shown in the figure, TCSNN adds a spiking neuron computing module on the basis of the temporal convolutional network TCN; the temporal convolutional network TCN contains C layers of residual network structures, and each layer of residual network structure is composed of two convolutional layers and one residual layer. The C layers of residual network structures are used to extract features and perform temporal modeling on the input feature sequence; the spiking neuron computing module includes X spiking neurons, and each spiking neuron is placed between two adjacent layers of residual network structures to play a connecting role. At the same time, based on the firing timing of the spiking neurons, the continuous signal output by the residual network structure is converted into a spiking signal and input to the spiking neurons to further optimize the prediction result;

[0129] The training process of TCSNN is as follows: First, initialize the hyperparameters and weights of TCSNN; then input the training set into TCSNN in batches for training until TCSNN converges; by adjusting the hyperparameters, obtain the optimal TCSNN; specifically as follows:

[0130] Step 2.1: Set the training batch size to 64, the learning rate to 0.001, the number of iterations to 300, and randomly initialize the neural weights of each layer of the TCSNN network model;

[0131] Step 2.2: Input the training set into TCSNN in batches and output the predicted value of the exhaust temperature of the aeroengine. Compare the predicted value of the exhaust temperature with the actual value, calculate the loss value Loss through the mean squared error function (MSE), and calculate the loss function for the weights of the convolutional layer and the residual layer using the chain rule, and update the weights of each convolutional layer and residual layer of TCSNN by backpropagation using the gradient descent algorithm;

[0132] Furthermore, the specific formula for the loss value Loss is:

[0133]

[0134] where is the number of samples, is the actual value, is the predicted value of TCSNN;

[0135] Furthermore, the gradient formula of the loss function for the weight is:

[0136]

[0137] Step 2.3: Judge whether the loss value is less than the threshold , when the loss value is less than the threshold When this happens, it is considered that the TCSNN has reached the optimum, and the training is stopped. Otherwise, the hyperparameters are adjusted, and step 2.2 is repeated;

[0138] Step 3: Evaluate the storage and computing resource requirements of the TCSNN on the FPGA; load the weight parameters obtained by training the TCSNN into the FPGA in the form of a COE file for storage;

[0139] Step 3.1: Select the Xilinx UltraScale+ series FPGA and configure 16-bit fixed-point numbers (Q2.13 format); the time convolution network calculation module adopts a 4-way parallel pipeline design, and each path is allocated 4 DSP48E2 units to achieve the calculation of a 4×3×3 convolution kernel in a single cycle; the ring RAM storage structure is as Figure 4 shown, with a depth of 2048, and a dual-port BRAM is used to achieve parallel reading and writing; the weight ROM adopts a segmented storage strategy, mapping the shallow weights with high-frequency access to the Block RAM, and the deep weights are stored in the distributed RAM;

[0140] Step 3.2: Write the weight parameters obtained by training the TCSNN into the COE file for the weight ROM module on the FPGA to read and use for the forward propagation of the TCSNN.

[0141] Step 4: The sensor module collects the parameters of each gas path component of the aeroengine in real time and inputs them into the FPGA, and compares the true value with the TCSNN prediction value to verify the network accuracy; draw the time series comparison chart of the true value and the TCSNN prediction value, and draw the comparison chart of the true value and the TCSNN prediction value as Figure 5 shown, the X-axis is the flight cycle, and the Y-axis is the absolute error between the true value and the TCSNN prediction value; the lower right corner is the comparison chart of the true value and the TCSNN prediction value, the X-axis is the TCSNN prediction value, and the Y-axis is the true value of this parameter. When X = Y, it means that there is no error between the TCSNN inference value and the true value.

Claims

1. A TCSNN-based aircraft engine gas path component degradation prediction system, characterized in that: The TCSNN-based aircraft engine gas path component degradation prediction system includes a data input module, a performance prediction module and a data output module deployed on the FPGA; The data input module includes a sensor module, a UART serial port receiving module, a data processing module, and a line buffer RAM module; the composition and specific functions of each module are as follows: Sensor module: including various types of sensors, used to collect raw data of the parameters of each gas path component changing over time; UART serial port receiving module: used to receive the raw data of the parameters of each gas path component changing over time collected by the sensor module, and generate characteristic data after parsing according to the UART protocol; Data processing module: used to process the characteristic data generated by the UART serial port receiving module according to the UART protocol analysis to obtain a characteristic sequence; The data processing module is deployed on the PS side of the FPGA, and the processed feature sequence is transmitted from the PS side to the PL side through the AXI bus; Row cache RAM module: used to cache the feature sequence processed by the data processing module; The performance prediction module includes a weight loading module, a weight ROM module, a time convolution network calculation module, a data flow control module, a pulse neuron calculation module, and a fully connected calculation module; the composition and specific functions of each module are as follows: Weight loading module: The temporal convolution network calculation module sends the weight address index to the weight loading module. The weight address index includes the convolution cycle count signal within a single layer and the overall convolution layer cycle count signal. The weight loading module generates the address of the weight parameter to be read according to the weight address index; Weight ROM module: used to write the weight parameters obtained by TCSNN training into the BRAM of FPGA in the form of COE file, and output the required weight parameters to the temporal convolutional network calculation module according to the address of the weight parameters generated by the weight loading module; Temporal convolution network computing module: including input data cache RAM unit, convolution calculation result cache RAM unit and convolution and residual general computing unit; Input data cache RAM unit: used to cache the characteristic sequence from the row cache RAM module in the data input module; Convolution calculation result cache RAM unit: used to cache the intermediate result data after the convolution and residual general calculation unit performs convolution calculation and residual calculation; Convolution and residual general computing unit: The convolution calculation parallelism is designed according to the shape of TCSNN and the DSP and LUT resources of FPGA. The convolution and residual general computing unit reads the feature sequence in the input data cache RAM unit and the intermediate result data in the convolution calculation result cache RAM unit according to the designed convolution calculation parallelism, and reads the weight parameters in the weight ROM module that matches the feature sequence to perform convolution calculation and residual calculation; during the convolution calculation and residual calculation process, the convolution calculation process of each layer needs to continuously update the convolution cycle count signal within the single layer to refresh the weight parameters; after the convolution calculation of each layer is completed, the overall cycle count signal of the convolution layer is updated, and The intermediate result data of the convolution calculation of the current layer is cached in the convolution calculation result cache RAM unit. When the convolution calculation of the next layer starts, the intermediate result data of the convolution calculation of the previous layer is directly read from the convolution calculation result cache RAM unit to perform the convolution calculation of the next layer. When the convolution calculation of all layers is completed, the convolution and residual general calculation unit performs residual calculation: read the feature sequence from the input data cache RAM unit, perform convolution calculation on the feature sequence and the refreshed weight parameter, and perform weighted addition on the result of the convolution calculation and the convolution calculation result of the last layer that has been cached in the convolution calculation result cache RAM unit. At this point, all calculations of the entire time convolution network calculation module are completed. Pulse neuron calculation module: The pulse neuron calculation module includes Z pulse neurons, according to the time step parameters designed for each pulse neuron Build Each pipeline structure is constructed by a group of LIF model operators. The membrane potential of the initial cache in the first pipeline structure is set to zero, and the membrane potential is transmitted to the next pipeline structure or an output pulse is generated according to the result signal output by the time convolution network calculation module. The pulse neuron calculation module converts the result signal output by the time convolution network calculation module into a pulse signal based on the firing sequence of the pulse neurons; the LIF model operator is explained by the following formula: （1） （2） Where: represents the time constant of the membrane, represents the membrane potential of the spiking neuron, represents the resting potential of a spiking neuron, represents the membrane resistance, Represents the external input current over time The formula (1) describes the influence of the membrane potential on the leakage effect of the resting potential and the external input current. represents the time when the spiking neuron fires a pulse, Indicates the time it takes for a spike neuron to approach 0 after it fires a spike The membrane potential of the endoplasmic neuron, represents the reset value of the membrane potential after the spiking neuron fires a pulse; Formula (2) describes the After the pulse is released at the moment, it will approach 0 reset its membrane potential to ; Data flow control module: The pulse signal output by the pulse neuron calculation module needs to be repeatedly input into the time convolution network calculation module C times for convolution calculation and residual calculation, and the pulse neuron calculation module is converted into a pulse signal. When the last time convolution network calculation module is completed, it is converted into a pulse signal by the pulse neuron calculation module and then handed over to the fully connected calculation module for calculation; The data flow control module controls the output flow between each module in the performance prediction module according to the convolution cycle count signal within the single layer and the convolution layer overall cycle count signal; Fully connected computing module: The fully connected computing module is responsible for mapping the pulse signal obtained by the pulse neuron computing module for the last time to the final output space, and maps the pulse signal to the predicted value of the exhaust temperature through a two-layer fully connected network; The data output module includes a UART serial port sending module and a threshold judgment module; UART serial port sending module: used to package the predicted value of exhaust temperature obtained by the performance prediction module according to the UART protocol and send it to the host computer through the serial port to display the predicted value of exhaust temperature and safety alarm; Threshold judgment module: used to monitor whether the predicted value of exhaust temperature is higher than the safety threshold. If the predicted value of exhaust temperature is higher than the safety threshold, a safety alarm will be issued.

2. The TCSNN-based aircraft engine gas path component degradation prediction system according to claim 1, characterized in that: Before deploying the TCSNN-based aircraft engine gas path component degradation prediction system on FPGA, TCSNN needs to be trained. The process is as follows: Step 1: collect the raw data of the parameters of each gas path component of the aircraft engine in cruise mode over time through sensors; preprocess the raw data, including normalization, noise reduction and data enhancement processing, to obtain the preprocessed gas path performance parameters; perform correlation analysis on each gas path performance parameter with the exhaust temperature, and use each gas path performance parameter with a correlation function value greater than a set threshold as a feature sequence input to TCSNN; Divide the feature sequence into training set, validation set and test set in proportion; The step 1 specifically comprises the following steps: Step 1.1: The raw data of the low-pressure rotor speed, high-pressure rotor speed, high-pressure compressor outlet pressure, oil pressure, compressor outlet temperature, and lead angle of the aircraft engine in cruise mode are collected with a fixed sampling frequency through sensors; Step 1.2: Use normalization to scale the raw data to The data sequence between; the wavelet packet threshold denoising method is used to denoise the normalized data sequence to obtain a denoised data sequence; the sliding overlapping sampling method is used to enhance the denoised data sequence; Step 1.2 specifically includes the following steps: Step 1.2.1: Normalize the original data and scale it to The number of samples is Data series ; Data Series Each of The calculation formula is as follows: in, is the original data, is the sequence of original data, is the maximum value in the sequence of original data, is the minimum value in the sequence of original data; Step 1.2.2, use The wavelet basis is used to normalize the data sequence. Layer wavelet decomposition, we get Layer wavelet coefficients, where each layer of wavelet decomposition results in approximate coefficients and detail coefficients; The detail coefficients of the layer are screened using a soft threshold function; The approximate coefficients of the layers and the soft threshold function are filtered from the first layer to the The detail coefficients of the layer are inversely transformed by wavelet transform to reconstruct the denoised data sequence ;in, for Data after noise reduction; Step 1.2.3: Use sliding overlapping sampling method and fixed sample length and overlap length The denoised data sequence is Split into A sample sequence, recorded as: in, For the denoised data series Middle The denoised data of the item; Step 1.3: Calculate the mutual information function value between each gas path performance parameter and the exhaust temperature after processing in step 1.

2. When the mutual information function value is greater than the set threshold, it is used as the feature sequence of the input TCSNN. When the mutual information function value is less than the set threshold, it is not used as the feature sequence of the TCSNN. Mutual information function between each gas path performance parameter and exhaust temperature The expression is: In the formula, is the denoised data sequence, Exhaust temperature series , , and They are , and The entropy of , and: In the formula, express The probability density of express The probability density of for and The joint probability density of Step 1.4: Divide the feature sequence obtained in step 1.3 into training set, validation set and test set in proportion; Step 2: Construction and training of the temporal convolutional spike neural network TCSNN; TCSNN adds a pulse neuron calculation module on the basis of the temporal convolutional network TCN; the temporal convolutional network TCN contains a C-layer residual network structure, each layer of the residual network structure consists of two convolutional layers and one residual layer, and the C-layer residual network structure is used to extract features and perform temporal modeling on the input feature sequence; the pulse neuron calculation module includes X pulse neurons, each of which is placed between two adjacent layers of residual network structures to play a connecting role. At the same time, based on the firing timing of the pulse neurons, the continuous signal output by the residual network structure is converted into a pulse signal input and transmitted to the pulse neuron, further optimizing the prediction results; The training process of TCSNN is as follows: first, initialize the hyperparameters and weights of TCSNN; then input the training set into TCSNN in batches for training until TCSNN converges; and obtain the optimal TCSNN by adjusting the hyperparameters; the details are as follows: Step 2.1: Initialize the hyperparameters and weights of TCSNN; Step 2.2: Input the training set into TCSNN in batches and output the predicted value of the exhaust temperature of the aircraft engine. Compare the predicted value of the exhaust temperature with the actual value, calculate the loss value Loss through the mean square error function, and calculate the loss function through the chain rule. Weights for convolutional and residual layers The gradient of is used to back-propagate and update the weights of each convolutional layer and residual layer of TCSNN; The specific formula of loss value Loss is: in, is the number of samples, is the actual value, is the predicted value of TCSNN; Loss Function Weight The gradient formula is: Step 2.3: Determine the loss value Is it less than the threshold? , when the loss value Less than threshold When , TCSNN is considered to be optimal and training is stopped. Otherwise, the hyperparameters are adjusted and step 2.2 is repeated. Step 3: Evaluate the storage and computing resource requirements of TCSNN on FPGA; load the weight parameters obtained by training TCSNN into FPGA in COE file format for storage; Step 3.1: Evaluate the usage of storage resources and logical computing resources for TCSNN and deployment based on the data format of weight parameters used in FPGA deployment, where on-chip DSP resources are used to calculate tensor multiplication in convolution and matrix calculation, and LUT is used to build the basic unit of TCSNN; Step 3.2: Write the weight parameters obtained by TCSNN training into the COE file for reading by the weight ROM module on the FPGA for forward propagation of TCSNN.

Images