Aero-engine blade manufacturing quality online evaluation method based on data analysis

Abstract

The application relates to the technical field of aviation manufacturing quality detection, and discloses an aviation engine blade manufacturing quality online evaluation method based on data analysis, wherein the method comprises the following steps: acquiring multi-channel process parameter time sequence data during the execution of each manufacturing procedure of a blade; performing timestamp alignment processing and amplitude normalization processing on each channel time sequence data to generate a synchronous process parameter multi-channel time sequence matrix; performing sliding window segmentation on the synchronous process parameter multi-channel time sequence matrix along a time axis; performing short-time Fourier transform on each time period sub-matrix to acquire time-frequency feature representation of each channel, and calculating the cross-correlation time-frequency feature matrix between each channel; and splicing the time-frequency features of each channel and the cross-correlation time-frequency features between channels along a feature dimension to generate a process fluctuation joint feature tensor of each time period.

Description

Technical Field

[0001] This invention relates to the field of aviation manufacturing quality inspection technology, specifically to an online evaluation method for the manufacturing quality of aero-engine blades based on data analysis. Background Technology

[0002] The manufacturing process of aero-engine blades involves multiple steps, including casting, forging, machining, and heat treatment. Process parameters such as temperature, pressure, vibration, and material flow rate in each step directly affect the microstructure and stress distribution within the blade.

[0003] The existing online quality assessment system adopts a "post-manufacturing inspection" working mode. In this mode, after the blade manufacturing process is completed, finished product data is collected through surface optical inspection or ultrasonic / X-ray non-destructive testing. This data is then compared with a quality standard database to output a pass / fail judgment. The process parameter time-series data recorded by the manufacturing execution system is only stored in the process log for post-event traceability and does not participate in the quality assessment calculation and decision-making process. The process parameter data channel and the quality inspection data channel are independent at the system architecture level.

[0004] The aforementioned architecture suffers from significant timeliness bottlenecks when handling internal blade defects. Subsurface microcracks, abnormal grain structure, and uneven residual stress distribution are internal defects that cannot be directly detected using surface optical inspection methods. The system must schedule ultrasonic or X-ray inspection processes, and quality assessment conclusions can only be output after non-destructive testing data acquisition and analysis are completed, leading to a prolonged online assessment response cycle. Furthermore, the formation of internal blade defects is strongly correlated with specific fluctuation patterns of process parameters during manufacturing. However, existing systems process process parameter time-series data and quality inspection data separately. The defect prediction information contained in the process parameters cannot be utilized in the quality assessment process. The system can only respond with an assessment after the defect has formed and been confirmed through time-consuming inspection, resulting in insufficient timeliness of online assessments. Summary of the Invention

[0005] This invention provides an online evaluation method for the manufacturing quality of aero-engine blades based on data analysis, which solves the technical problem in related technologies that cannot simultaneously integrate the dynamic characteristics of process parameters and surface inspection information during the manufacturing process, resulting in insufficient real-time evaluation of blade quality and limited defect location capabilities.

[0006] This invention discloses an online evaluation method for the manufacturing quality of aero-engine blades based on data analysis, comprising the following steps: during the execution of each manufacturing process of the blade, acquiring multi-channel process parameter time-series data; performing timestamp alignment and amplitude normalization processing on the time-series data of each channel to generate a synchronous process parameter multi-channel time-series matrix; performing sliding window segmentation along the time axis on the synchronous process parameter multi-channel time-series matrix; performing short-time Fourier transform on the sub-matrices of each time period to obtain the time-frequency feature representation of each channel; calculating the cross-correlation time-frequency feature matrix between each pair of channels; concatenating the time-frequency features of each channel itself and the cross-correlation time-frequency features between channels along the feature dimension to generate a joint feature tensor of process fluctuations for each time period; and combining multiple... The process fluctuation joint feature tensor is sequentially input into a bidirectional long short-term memory network to obtain a hidden state vector sequence. After weighted aggregation by a self-attention calculation module, a global encoding vector of the process is output. Fast surface detection data of the same blade is obtained, and observable surface quality feature vectors are extracted by a convolutional neural network. The global encoding vector of the process and the observable surface quality feature vectors are input into a cross-attention fusion module to generate a collaborative fusion representation vector. The collaborative fusion representation vector is input into a multi-label prediction network to output the existence probability value and severity level code of each defect type. The comprehensive quality risk index of the blade is calculated, and an evaluation conclusion is output based on the comparison result of the comprehensive quality risk index and a preset safety threshold.

[0007] Furthermore, the cross-correlation time-frequency characteristic matrix is ​​calculated as follows: Frequency-wise conjugate product operations are performed on the frequency component vectors of the two channels in the same time frame to obtain the cross-correlation spectrum vector of the corresponding time frame; the cross-correlation spectrum vectors of each time frame are arranged in time frame order to form a matrix; assuming the two channels are in the [missing information - likely a specific time frame], the cross-correlation spectrum vectors of each time frame are [missing information - likely a specific time frame]. The frequency component vectors in each time frame are respectively and ,in For time frame indexing, Given the total number of frequency components, the cross-correlation spectrum vector of that time frame is... The The elements are ,in For frequency component index, for The conjugate of complex numbers.

[0008] Furthermore, the multi-channel process parameter time-series data covers temperature field distribution sequence, pressure change sequence, vibration spectrum sequence, and material flow rate sequence; there are six cross-correlation time-frequency feature matrices between each pair of channels, corresponding to six channel pairs; the process fluctuation joint feature tensor is formed by splicing the time-frequency feature representation matrices of the four channels themselves and the six cross-correlation time-frequency feature matrices along the feature dimension.

[0009] Furthermore, the calculation process of the self-attention calculation module is as follows: the hidden state vectors of each time period are mapped to query vectors, key vectors and value vectors through three independent linear transformation layers; the scaled dot product of the query vector and the key vector is calculated to obtain the attention weight matrix; the attention weight matrix is ​​multiplied by the value vector to obtain the weighted features; the weighted features are summed along the time period dimension to obtain the global encoding vector of the process; wherein, the self-attention calculation module adopts a multi-head attention form, splits the hidden state vector into multiple subspaces, each subspace independently calculates the attention weight, and the weighted outputs of each subspace are concatenated and mapped back to the original dimension through a linear transformation layer.

[0010] Furthermore, the fast surface detection data includes surface roughness distribution map data and surface morphology grayscale image data; before inputting into the convolutional neural network, Z-score normalization is performed on the surface roughness distribution map data and the surface morphology grayscale image data respectively, and their respective pixel values ​​are mapped to a unified dimensionless distribution and then concatenated along the channel dimension to form a multi-channel input.

[0011] Furthermore, in the cross-attention fusion module, the global encoding vector of the process is mapped to a query matrix through a linear transformation layer. The surface observable quality feature vector is mapped to a bond matrix through two independent linear transformation layers. AND-value matrix The formula for calculating scaled dot product attention is: in, For querying the matrix, The key matrix, For value matrices, Indicates transpose. Let be the dimension of the key vector. It is a normalized exponential function.

[0012] Furthermore, the comprehensive quality risk index of the blade The calculation formula is: in, This represents the total number of defect types. For defect type index, For the first Preset weighting coefficients for each defect type For the first The probability value of the existence of each defect type For the first Severity level coding for different defect types; when Less than the preset safety threshold When a qualified assessment conclusion is generated, the blade skips the subsequent non-destructive testing process; when Greater than or equal to At that time, an evaluation conclusion requiring detailed non-destructive testing is generated.

[0013] Furthermore, when Greater than or equal to In addition, a set of recommended detection area coordinates and a code for recommended detection method are output. The set of recommended detection area coordinates is generated by extracting the process stage identifiers corresponding to several time periods with the highest attention weight in the self-attention calculation module, and determining the spatial coordinate range of the region with the most drastic fluctuations in process parameters on the blade based on the correspondence between the process stage identifiers and the blade geometric model. The code for recommended detection method is determined based on the defect type with the highest probability of existence output by the multi-label prediction network.

[0014] Furthermore, when calculating the comprehensive quality risk index of the blade, a nonlinear mapping is performed on the probability values ​​of each defect type: let the probability warning threshold be... For the first Types of defects, when At that time, with replace Participating in the weighted summation operation, among which Represents an exponential function; when hour, It directly participates in the weighted summation operation.

[0015] This invention discloses an online evaluation system for the manufacturing quality of aero-engine blades, used to execute the aforementioned online evaluation method for the manufacturing quality of aero-engine blades. The system includes: a process data acquisition and preprocessing module, used to acquire multi-channel process parameter time-series data during the execution of each manufacturing process of the blade, perform timestamp alignment and amplitude normalization processing on each channel's time-series data, and generate a synchronous process parameter multi-channel time-series matrix; a process fluctuation feature extraction module, used to perform sliding window segmentation on the synchronous process parameter multi-channel time-series matrix, perform short-time Fourier transform and inter-channel cross-correlation calculation on each time-segment submatrix, and generate a process fluctuation joint feature tensor for each time segment; and a process global encoding module, used to sequentially input multiple process fluctuation joint feature tensors into bidirectional long and short time-varying arrays. A time-memory network, weighted and aggregated by a self-attention calculation module, outputs a global encoding vector for the process. A surface feature extraction module acquires rapid surface inspection data of the blade, which, after standardization and concatenation with the channel dimension, is input into a convolutional neural network to output a surface observable quality feature vector. A collaborative fusion module inputs the global encoding vector for the process and the surface observable quality feature vector into a cross-attention fusion module to generate a collaborative fusion representation vector. A quality assessment output module inputs the collaborative fusion representation vector into a multi-label prediction network to output the probability value of each defect type and the severity level code, calculates the blade's comprehensive quality risk index, and outputs an assessment conclusion based on the comparison result of the comprehensive quality risk index and a preset safety threshold.

[0016] The present invention has the following beneficial effects.

[0017] This invention transforms signals related to the formation of internal defects, such as abnormal temperature gradients and pressure fluctuations, from the time domain to a joint feature representation in the time and frequency domain by performing short-time Fourier transform and cross-correlation calculation on the time-series data of multi-channel process parameters. This makes the defect correlation information contained in the process parameters computable, overcoming the limitation of existing systems where process parameter data is only used for post-event traceability and cannot participate in quality assessment calculations.

[0018] This invention encodes the joint feature tensor of process fluctuations scattered across different time periods into a global encoding vector of the process by using a bidirectional long short-term memory network and self-attention weighted aggregation. This preserves the forward and backward temporal dependencies of process parameters throughout the entire manufacturing process, enabling the global encoding vector to reflect the process fluctuation patterns across time periods during the manufacturing process.

[0019] This invention uses a cross-attention fusion module with the global encoding vector of the process as the query driver to perform association retrieval on the surface observable quality feature vector. It establishes a mapping relationship between process anomaly patterns and surface observable anomalies in the feature space, enabling the process parameter data channel and the quality inspection data channel to be fused at the feature level, overcoming the architectural limitation of the two types of data channels being independent in the existing system.

[0020] This invention directly calculates the comprehensive quality risk index based on collaborative fusion representation vectors. For blades with a comprehensive quality risk index lower than a preset safety threshold, it generates a qualified assessment conclusion and skips the time-consuming non-destructive testing process. It only triggers targeted non-destructive testing for high-risk blades, thereby decoupling the online assessment response cycle from the full non-destructive testing cycle and alleviating the problem of insufficient online assessment timeliness caused by the separation of process data and testing data in existing systems. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] Example 1 The manufacturing process of aero-engine blades involves multiple steps, including casting, forging, machining, and heat treatment. Process parameters such as temperature, pressure, vibration, and material flow rate in each step directly affect the microstructure and stress distribution within the blade. Existing online quality assessment systems employ a "post-manufacturing inspection" model. This means that after the blade manufacturing process is complete, finished product data is collected through surface optical inspection or ultrasonic / X-ray non-destructive testing, and then compared with a quality standard database to output a pass / fail judgment. In this system architecture, the time-series data of process parameters recorded by the manufacturing execution system is only stored in the process log for post-processing traceability and does not participate in the quality assessment calculation and decision-making process. The process parameter data channel and the quality inspection data channel are independent at the system architecture level.

[0023] The aforementioned architecture suffers from significant time-sensitivity bottlenecks when handling internal blade defects (including subsurface microcracks, abnormal grain structure, and uneven residual stress distribution). Internal defects cannot be directly detected using surface optical inspection methods; the system must schedule ultrasonic or X-ray inspection processes, and quality assessment conclusions can only be output after non-destructive testing data acquisition and analysis, leading to a prolonged online assessment response cycle. However, the formation of internal blade defects is strongly correlated with specific fluctuation patterns of process parameters during manufacturing. For example, abnormal temperature gradient sequences at specific times during the casting process are highly correlated with the probability of shrinkage cavities. Because existing systems process process parameter time-series data and quality inspection data separately, the defect prediction information contained in the process parameters cannot be utilized in the quality assessment process; the system can only provide an assessment response after the defect has formed and been confirmed by time-consuming inspection.

[0024] To address the aforementioned issues, this invention proposes an online evaluation method for the manufacturing quality of aero-engine blades based on data analysis. By transforming the time-series data of process parameters into a computable time-frequency joint feature representation and synergistically fusing it with rapid surface inspection data in the feature space, the method enables the prediction of internal defects in the blades, thereby decoupling the online evaluation response cycle from the full-scale non-destructive testing cycle.

[0025] like Figure 1 As shown, an online evaluation method for the manufacturing quality of aero-engine blades based on data analysis according to an embodiment of the present invention includes the following steps: Step 1: Acquire and preprocess multi-channel process parameter timing data.

[0026] During each manufacturing process of the blade, multi-channel time-series data of process parameters are acquired in real time at a preset sampling frequency. The data channels cover temperature field distribution sequences, pressure change sequences, vibration spectrum sequences, and material flow rate sequences. Timestamp alignment is performed on the time-series data of each channel to ensure that the data of each channel has a unified sampling time on the time axis. Amplitude normalization is then performed on the timestamp-aligned data of each channel to eliminate the magnitude differences between different physical dimensions. After the above processing, a synchronous process parameter multi-channel time-series matrix is ​​generated and stored in the real-time data buffer.

[0027] It should be noted that the synchronous process parameter multi-channel time series matrix is ​​a two-dimensional matrix, with rows corresponding to each sampling time and columns corresponding to each process parameter channel. Timestamp alignment is performed using the timestamp of the channel with the lowest sampling frequency as a reference, and a linear interpolation algorithm is applied to the data of the remaining channels to ensure that all channels have corresponding values ​​on the same time scale. Amplitude normalization is performed on the data of each channel, using a mean normalization based on the range of the maximum and minimum values ​​within the current process time range, mapping the channel values ​​to the interval between zero and one. After normalization, the values ​​of the temperature field distribution sequence, pressure change sequence, vibration spectrum sequence, and material flow rate sequence are all within the same dimensionless interval, allowing them to be directly used in subsequent matrix operations and feature extraction.

[0028] Step 2: Extract the joint feature tensor of process fluctuations.

[0029] The synchronous process parameter multi-channel timing matrix is ​​divided along the time axis using a preset window length and preset step size, generating multiple time-segment sub-matrices. Each time-segment sub-matrix contains the process parameter values ​​of all channels within the corresponding time window.

[0030] Short-time Fourier transforms (SFTs) are calculated for each channel data in each time-segment submatrix to obtain the time-frequency characteristic representation of each channel. The time-frequency characteristic representation of each channel is a two-dimensional matrix, where rows correspond to frequency components and columns correspond to time frames within the window. The SFTs of the four channels—temperature field distribution sequence, pressure change sequence, vibration spectrum sequence, and material flow rate sequence—are calculated independently, yielding corresponding time-frequency characteristic representation matrices, for a total of four.

[0031] Further calculations are performed on the cross-correlation time-frequency characteristic matrices between each pair of channels. The cross-correlation time-frequency characteristic matrices characterize the joint fluctuation relationship between the two channels at each frequency component.

[0032] The time-frequency characteristics of each channel are concatenated with the corresponding cross-correlation time-frequency characteristics between channels along the feature dimension to generate a joint feature tensor of process fluctuations for each time period.

[0033] It should be noted that the cross-correlation time-frequency characteristic matrix is ​​calculated as follows: For the frequency component vectors of the two channels in the same time frame, perform a frequency-wise conjugate product operation to obtain the cross-correlation spectrum vector of the corresponding time frame. Then, arrange the cross-correlation spectrum vectors of each time frame in time frame order to form a matrix. Specifically, let the two channels be in the... The frequency component vectors in each time frame are respectively and ,in For time frame indexing, Given the total number of frequency components in the short-time Fourier transform, the cross-correlation spectrum vector of that time frame is... The The elements are ,in For frequency component index, , for The conjugate complex number. The cross-correlation spectrum vectors of each time frame. Arranged in time frame order, they form a cross-correlation time-frequency feature matrix, where This represents the total number of time frames within the window. For the four process parameter channels, the pairwise channel pairs include: temperature field distribution sequence and pressure change sequence, temperature field distribution sequence and vibration spectrum sequence, temperature field distribution sequence and material flow rate sequence, pressure change sequence and vibration spectrum sequence, pressure change sequence and material flow rate sequence, and vibration spectrum sequence and material flow rate sequence, totaling six pairs. The cross-correlation time-frequency feature matrix of each pair is calculated as described above, resulting in six cross-correlation time-frequency feature matrices. The four time-frequency feature representation matrices and the six cross-correlation time-frequency feature matrices are concatenated along the feature dimensions to form the joint feature tensor of process fluctuations for that time period. The joint feature tensor of process fluctuations simultaneously encodes the frequency domain fluctuation characteristics of each channel and the cooperative fluctuation relationship between channels.

[0034] Step 3: Generate the global encoding vector for the process.

[0035] Multiple joint feature tensors of process fluctuations, arranged in chronological order of process operation time, are sequentially input into a bidirectional long short-term memory (LSTM) network. The input layer of the bidirectional LSM network receives a flattened one-dimensional vector sequence of the joint feature tensors of process fluctuations for each time period. The output layer outputs a sequence of hidden state vectors corresponding to each time period. The hidden state vector for each time period is formed by concatenating the forward and backward hidden states, thereby fusing the forward and backward temporal dependencies and obtaining the global temporal evolution features of process parameter fluctuations.

[0036] The global temporal evolution features of process parameter fluctuations are input into the self-attention calculation module. The self-attention calculation module takes the hidden state vector sequence of each time period as input and calculates the attention weight distribution among the features of each time period. The attention weight represents the contribution of each time period to the quality assessment objective. Weighted aggregation is performed on the global temporal evolution features of process parameter fluctuations according to the attention weights, outputting a global encoding vector of the process.

[0037] The bidirectional long short-term memory network and the self-attention computation module are jointly trained, with the multi-label classification loss of the multi-label prediction network in step 6 as the supervision signal, and end-to-end training is performed using the backpropagation algorithm and the Adam optimization algorithm.

[0038] It should be noted that the calculation process of the self-attention calculation module is as follows: the hidden state vectors of each time period are mapped to query vector, key vector and value vector through three independent linear transformation layers; the scaled dot product of query vector and key vector is calculated to obtain attention weight matrix; the attention weight matrix is ​​multiplied by value vector to obtain weighted features; the weighted features are summed along the time period dimension to obtain global encoding vector of process.

[0039] In this embodiment, to enhance the representation ability of the global process encoding vector for critical process anomalies, the attention weight calculation in the self-attention calculation module can adopt a multi-head attention approach. Multi-head attention splits the hidden state vector into multiple subspaces, each of which independently calculates its attention weight. Finally, the weighted outputs of each subspace are concatenated and mapped back to the original dimension via a linear transformation layer. Multi-head attention enables the global process encoding vector to simultaneously capture process fluctuation patterns at different scales.

[0040] Step 4: Extract the surface observable quality feature vector.

[0041] Acquire rapid surface inspection data for the same blade at the current process completion node. The rapid surface inspection data includes surface roughness distribution map data and surface morphology grayscale image data. The surface roughness distribution map data is a two-dimensional distribution map with each sampling point on the blade surface as the pixel location and the corresponding roughness value as the pixel value; the surface morphology grayscale image data is a grayscale image of the blade surface acquired by an optical camera. Since the pixel values ​​of the surface roughness distribution map data and the surface morphology grayscale image data have different dimensions, Z-score normalization is performed on both data before concatenation along the channel dimension. This maps their respective pixel values ​​to a unified dimensionless distribution, eliminating the impact of the dimensional difference on subsequent convolution operations. After normalization, the surface roughness distribution map data and the surface morphology grayscale image data are concatenated along the channel dimension to form a multi-channel input.

[0042] Multi-channel input data is fed into a convolutional neural network (CNN). The input layer of the CNN receives the multi-channel input data and extracts spatial features such as texture, undulations, and local anomalies from the surface data through multiple convolutional and pooling layers. The output layer is a fully connected layer that outputs a fixed-length surface observable quality feature vector. The CNN is trained end-to-end using the multi-label classification loss of the multi-label prediction network in step 6 as a supervision signal, and employs the backpropagation algorithm and the Adam optimization algorithm.

[0043] Step 5: Generate collaborative fusion representation vectors.

[0044] The global encoding vector of the process and the surface observable quality feature vector are input into the cross-attention fusion module, and scaled dot product attention calculation is performed to generate a collaborative fusion representation vector.

[0045] It should be noted that in the cross-attention fusion module, the global encoding vector of the process is mapped to a query matrix through a linear transformation layer. The observable quality eigenvectors of the surface are mapped to bond matrices through two independent linear transformation layers. AND-value matrix ,in , , All column dimensions are , The feature dimension for attention calculation. The formula for scaling dot product attention is: in, For querying the matrix, The key matrix, For value matrices, Indicates transpose. Let be the dimension of the key vector. This is a normalized exponential function. The global encoding vector of the process and the surface observable quality feature vector have both been mapped to a column dimension of [value missing] before being input into the cross-attention fusion module, respectively, through their respective linear transformation layers. The unified feature space, therefore the query matrix AND key matrix With consistent column dimensions, efficient dot product operations can be performed. The scaled dot product attention computation uses the global encoding vector of the process as the query driver, performs correlation retrieval on the surface observable quality feature vector, and establishes a mapping relationship between process anomaly patterns and surface observable anomalies in the feature space, enabling the fusion of two originally independent data channels at the feature level.

[0046] Step 6: Calculate the comprehensive quality risk index and output the assessment conclusion.

[0047] The co-fusion representation vector is input into the multi-label prediction network. The multi-label prediction network is a fully connected neural network. Its input layer receives the co-fusion representation vector, and after processing through multiple fully connected layers, the output layer contains two parallel branches: an existence probability output branch and a severity level output branch. The existence probability output branch corresponds to... Each output node is activated by a sigmoid function, and the output node outputs the probability value of the existence of the corresponding defect type. ,in For defect type index, The value range is from zero to one; the severity level output branch also corresponds to this. Each output node outputs a severity level code corresponding to a defect type. ,in The values ​​are continuous real numbers, converted to discrete integer values ​​through rounding and mapping before being used in subsequent calculations. Larger values ​​indicate a higher estimated severity. Defect types include subsurface microcracks, shrinkage cavities, abnormal grain structure, and uneven residual stress distribution, totaling [number missing]. kind, The total number of defect types is represented by . The multi-label prediction network uses a binary cross-entropy loss function, with the true labels of each defect type as the supervision signal, and is trained using the backpropagation algorithm and the Adam optimization algorithm.

[0048] The blade's overall quality risk index is calculated by performing a weighted summation operation on the probability values ​​of each defect type and the severity level codes according to preset weighting coefficients. The calculation formula is: in, This is the comprehensive quality risk index for blades. This represents the total number of defect types. For defect type index, For the first Preset weighting coefficients for each defect type For the first The probability value of the existence of each defect type For the first Severity level coding for each defect type. The value ranges from zero to one. For discrete integer values, For the pre-defined dimensionless weighting coefficients, the product of the three factors and their summation are consistent in dimensions, resulting in the comprehensive quality risk index. These are dimensionless values. The preset weighting coefficients for each defect type are pre-set based on the degree of impact of each type of defect on the service safety of the blade.

[0049] Comprehensive quality risk index With preset safety threshold Perform numerical comparisons. When At that time, a pass / fail assessment conclusion is generated and output, and the blade skips the subsequent time-consuming non-destructive testing process. When the time comes, an evaluation conclusion requiring detailed non-destructive testing is generated, along with an output of a recommended set of test area coordinates and a recommended test method code.

[0050] It should be noted that the recommended detection area coordinate set is generated as follows: In the self-attention calculation module of step 3, the process stage identifiers corresponding to the time periods with the highest attention weights are extracted; based on the correspondence between the process stage identifiers and the blade geometric model, the spatial coordinate range of the region with the most drastic fluctuations in process parameters during that process stage is determined, and this range is output as the recommended detection area coordinate set. The recommended detection method code is determined based on the defect type with the highest existence probability value output by the multi-label prediction network: for subsurface microcrack defects, the recommended detection method code corresponds to ultrasonic testing; for shrinkage cavities or abnormal grain structure defects, the recommended detection method code corresponds to X-ray testing.

[0051] In this embodiment of the application, to improve the sensitivity of the comprehensive quality risk index in distinguishing defects of different severity levels, the following step is included in addition to the weighted summation operation: performing a nonlinear mapping on the existence probability value of each defect type. Let the probability warning threshold be... For the first Types of defects, when their probability value exists At that time, with replace Participating in the weighted summation operation, among which Represents an exponential function; when hour, It directly participates in the weighted summation calculation. This processing makes the comprehensive quality risk index more sensitive to high-probability defects, avoiding the cumulative effect of multiple low-probability defects from masking the early warning signal of a single high-probability defect.

[0052] On the other hand, the present invention also proposes an online evaluation system for the manufacturing quality of aero-engine blades, comprising: The process data acquisition and preprocessing module is used to acquire multi-channel process parameter time-series data in real time at a preset sampling frequency during the execution of each manufacturing process of the blade, perform timestamp alignment and amplitude normalization processing on the time-series data of each channel, and generate a synchronous process parameter multi-channel time-series matrix.

[0053] The process fluctuation feature extraction module is used to perform sliding window segmentation on the multi-channel time series matrix of synchronous process parameters with a preset window length and preset step size, perform short-time Fourier transform and cross-correlation calculation on the sub-matrices of each time period, and generate the joint feature tensor of process fluctuation for each time period.

[0054] The process global encoding module is used to sequentially input multiple process fluctuation joint feature tensors into a bidirectional long short-term memory network and then weight and aggregate them through a self-attention calculation module to output a process global encoding vector.

[0055] The surface feature extraction module is used to acquire rapid surface detection data of the blade at the current process completion node. It performs Z-score standardization on the surface roughness distribution map data and the surface morphology grayscale image data respectively, then stitches them along the channel dimension, inputs them into the convolutional neural network, and outputs the surface observable quality feature vector.

[0056] The collaborative fusion module is used to input the global encoding vector of the process and the surface observable quality feature vector into the cross-attention fusion module, perform scaled dot product attention calculation, and generate a collaborative fusion representation vector.

[0057] The quality assessment output module is used to input the collaborative fusion representation vector into the multi-label prediction network, output the existence probability value and severity level code of each defect type, calculate the blade comprehensive quality risk index, and output the assessment conclusion based on the comparison result of the comprehensive quality risk index and the preset safety threshold.

[0058] By way of example, the system proposed in this invention may also include various features and combinations thereof in the method embodiments, which will not be elaborated here.

[0059] Technical effect This implementation method performs short-time Fourier transform and inter-channel cross-correlation calculations on the multi-channel process parameter time-series data in step 2. This transforms signals related to internal defect formation, such as temperature gradient anomalies and pressure fluctuations, from the time domain to a joint feature representation in the time-frequency domain, making the defect correlation information contained in the process parameters computable. Step 3 uses a bidirectional long short-time memory network and self-attention weighted aggregation to encode the joint feature tensor of process fluctuations scattered across different time periods into a global process encoding vector, preserving the forward and backward temporal dependencies of process parameters throughout the manufacturing process. Step 5, cross-attention fusion, uses the global process encoding vector as a query driver to perform correlation retrieval on the surface observable quality feature vector. This establishes a mapping relationship between process anomaly patterns and surface observable anomalies in the feature space, enabling the process parameter data channel and the quality inspection data channel to be integrated at the feature level, overcoming the architectural limitations of existing systems where the two types of data channels are independent. Step 6 directly calculates the comprehensive quality risk index based on the collaborative fusion representation vector. For blades with a comprehensive quality risk index lower than the preset safety threshold, a qualified assessment conclusion is generated and the time-consuming non-destructive testing process is skipped. Targeted non-destructive testing is triggered only for high-risk blades, thereby decoupling the online assessment response cycle from the full non-destructive testing cycle and solving the bottleneck of online assessment timeliness caused by the separation of process data and testing data in the existing system.

[0060] An aircraft manufacturing plant performed an online quality assessment on blade B-0471 during the casting process of turbine guide vanes. This blade was cast using a directionally solidified nickel-based superalloy, and the casting process lasted approximately 180 minutes. The process data acquisition system simultaneously recorded data from four channels at a sampling frequency of 50Hz: temperature field distribution sequence, pressure change sequence, vibration spectrum sequence, and material flow rate sequence. After the process was completed, optical inspection equipment simultaneously acquired surface roughness distribution maps and grayscale images of the surface morphology. The objective of this assessment was to make a real-time determination of potential internal defects in the blade without scheduling full-scale non-destructive testing.

[0061] During the casting process, the raw sampling data from the four channels exhibited significant dimensional differences. The temperature field distribution sequence ranged from approximately 1200 to 1480. The pressure variation sequence ranges from approximately 0.8 to 3.2 MPa, and the vibration spectrum sequence is characterized by the effective value of acceleration, ranging from approximately 0.12 to 2.85 MPa. The material flow rate sequence ranges from approximately 0.03 to 0.18 m / s. Timestamp alignment is performed using the material flow rate sequence with the lowest sampling frequency (10 Hz) as the base timestamp. Linear interpolation is then applied to the data from the other three channels to align all channels on the same time scale. After alignment, mean normalization based on the range is performed on each channel, mapping the values ​​to the zero-to-one interval. This generates a multi-channel time series matrix of synchronized process parameters, which is then stored in the real-time data buffer.

[0062] Table 1. Multi-channel timing matrix of synchronous process parameters (normalized, partial sampling times) A sliding window segmentation was performed along the time axis on the multi-channel time series matrix of synchronous process parameters with a preset window length of 30 minutes and a step size of 15 minutes, generating a total of 11 time-segment sub-matrices. Short-time Fourier transforms were calculated for each channel data in each time-segment sub-matrix to obtain the time-frequency feature representation matrix for each channel. Taking the fourth time segment (covering minutes 45 to 75) as an example, the temperature field distribution sequence exhibited an abnormal concentration of energy in the low-frequency range during this segment, corresponding to intensified temperature gradient fluctuations during the casting process. Based on this, six pairs of cross-correlation time-frequency feature matrices were calculated between the four channels. The cross-correlation time-frequency feature matrices of the temperature field distribution sequence and the pressure change sequence showed a significant positive correlation concentration in the low-frequency range (0 to 5 Hz), indicating that the two channels exhibited coordinated fluctuations in this frequency band. The four self-time-frequency feature matrices and the six cross-correlation time-frequency feature matrices were concatenated along the feature dimension to generate the joint feature tensor of process fluctuations for each time segment.

[0063] Table 2. Joint characteristic tensor description of process fluctuations in different time periods (partial time periods) The joint feature tensors of process fluctuations across 11 time periods were flattened and input into a bidirectional long short-term memory network in chronological order. The forward channel of the network captures the influence of the early temperature establishment stage of casting on the subsequent solidification process, while the backward channel captures the inverse correlation between pressure changes at the end of solidification and the early formed microstructure. The forward and backward hidden states output from each time period were concatenated and input into the self-attention calculation module. The attention weight calculation results show that the attention weights obtained for time periods 4 and 5 are significantly higher than those for other time periods, indicating that the process fluctuations from minute 45 to minute 90 contribute the most to the quality assessment objective. Weighted aggregation was performed on all hidden state vectors according to the attention weights to output a global encoding vector of the process.

[0064] Table 3. Distribution of attention weights in different time periods (partial time periods) After the process is completed, the optical inspection equipment collects surface roughness distribution data (pixel values ​​in μm) and surface morphology grayscale image data (pixel values ​​ranging from 0 to 255) of blade B-0471. Since the two types of data have different dimensions, Z-score normalization is performed on each, mapping their respective pixel values ​​to a dimensionless distribution with a mean of zero and a variance of one. After normalization, the two types of data are concatenated along the channel dimension to form a dual-channel input, which is then fed into a convolutional neural network. Through multiple layers of convolution and pooling operations, the network detects local texture anomaly concentration bands in the chordal region of the blade and outputs a fixed-length surface observable quality feature vector.

[0065] Table 4 Comparison of surface inspection data before and after standardization The global encoding vector of the process is mapped to a query matrix through a linear transformation layer. The observable quality eigenvectors of the surface are mapped to bond matrices through two independent linear transformation layers. AND-value matrix The three dimensions are unified as follows: The cross-attention fusion module is configured according to the formula. Perform scaling dot product attention calculation. Driven by the global encoding vector of the process, the module retrieves the correlation response between the texture anomaly in the mid-sine region and the temperature gradient fluctuation from the 45th to the 90th minute in the surface observable quality feature vector, establishes the mapping relationship between the process anomaly pattern and the surface observable anomaly in the feature space, and generates a collaborative fusion representation vector.

[0066] The collaborative fusion representation vector is input into the multi-label prediction network. The existence probability output branch outputs the existence probability value for each of the four defect types after passing through the Sigmoid activation function. The severity level output branch outputs the corresponding severity level code. For shrinkage cavity defects, the probability value of their existence is... Exceeding the probability warning threshold Therefore, replace By participating in the weighted summation calculation, the comprehensive quality risk index becomes more sensitive to this high-probability defect. According to the formula... Calculate the comprehensive quality risk index, where the item corresponding to shrinkage cavity defects is... Substitution Location is involved in the calculation, and a preset safety threshold is set. .

[0067] Table 5. Multi-label prediction network output and calculation of comprehensive quality risk index Comprehensive quality risk index The calculation result exceeds the preset safety threshold. The system generates an assessment conclusion that "detailed non-destructive testing is required." The time periods 4 and 5, with the highest attention weights in the attention calculation module, correspond to the temperature gradient fluctuation stage and the pressure vibration co-abnormal stage from minute 45 to minute 90 of the casting process. Based on the correspondence between this process stage and the blade geometry model, the recommended detection area coordinate set is located in the inner cavity region from the blade's center chord to the trailing edge. Since the most probable defect type is shrinkage cavity, the recommended detection method is coded as X-ray inspection. The system outputs the above conclusions and recommendations together.

[0068] Throughout the evaluation process, the data starts from the four-channel raw process parameter time-series data. After timestamp alignment and normalization, a synchronous process parameter multi-channel time-series matrix is ​​generated. This matrix is ​​then transformed into a process fluctuation joint feature tensor sequence through sliding window segmentation, short-time Fourier transform, and cross-correlation calculation. Finally, it is compressed into a global encoding vector of the process through a bidirectional long short-time memory network and self-attention weighted aggregation. Simultaneously, the surface inspection data is transformed into a surface observable quality feature vector through Z-score normalization and convolutional neural network processing. The two feature paths are collaboratively fused in the cross-attention fusion module with process anomaly patterns as the query driver. Finally, the multi-label prediction network outputs the defect probability and severity, and the threshold comparison of the comprehensive quality risk index directly triggers targeted detection conclusions. The entire process does not rely on the full amount of non-destructive testing data, thus decoupling the online evaluation response cycle from the full amount of non-destructive testing cycle.

[0069] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A data analysis-based online evaluation method for the manufacturing quality of aero-engine blades, characterized in that, Includes the following steps: During the execution of each manufacturing process of the blade, multi-channel process parameter timing data is acquired, and timestamp alignment and amplitude normalization are performed on the timing data of each channel to generate a synchronous process parameter multi-channel timing matrix. The multi-channel timing matrix of the synchronous process parameters is divided into sliding window segments along the time axis. Short-time Fourier transform is performed on the sub-matrix of each time period to obtain the time-frequency feature representation of each channel. The cross-correlation time-frequency feature matrix between each pair of channels is calculated. The time-frequency features of each channel itself and the cross-correlation time-frequency features between channels are concatenated along the feature dimension to generate the joint feature tensor of process fluctuations for each time period. The joint feature tensors of multiple process fluctuations are sequentially input into a bidirectional long short-term memory network to obtain a sequence of hidden state vectors. After weighted aggregation by the self-attention calculation module, the global encoding vector of the process is output. Acquire rapid surface detection data of the same blade, and extract observable surface quality feature vectors using a convolutional neural network; The global encoding vector of the process and the observable quality feature vector of the surface are input into the cross-attention fusion module to generate a collaborative fusion representation vector. The collaborative fusion representation vector is input into a multi-label prediction network, which outputs the probability value of the existence of each defect type and the severity level code. The comprehensive quality risk index of the blade is calculated, and the evaluation conclusion is output based on the comparison result of the comprehensive quality risk index and the preset safety threshold.

2. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 1, characterized in that, The cross-correlation time-frequency characteristic matrix is ​​calculated as follows: Frequency-wise conjugate product operations are performed on the frequency component vectors of the two channels in the same time frame to obtain the cross-correlation spectrum vector of the corresponding time frame. The cross-correlation spectrum vectors of each time frame are then arranged in time frame order to form a matrix. Let the two channels be in the [missing information - likely a specific time frame]. The frequency component vectors in each time frame are respectively and ,in For time frame indexing, Given the total number of frequency components, the cross-correlation spectrum vector of that time frame is... The The elements are ,in For frequency component index, for The conjugate of complex numbers.

3. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 1, characterized in that, The multi-channel process parameter time-series data covers temperature field distribution sequence, pressure change sequence, vibration spectrum sequence, and material flow rate sequence; there are six cross-correlation time-frequency feature matrices between each pair of channels, corresponding to six channel pairs; the process fluctuation joint feature tensor is formed by splicing the time-frequency feature representation matrices of the four channels themselves and the six cross-correlation time-frequency feature matrices along the feature dimension.

4. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 1, characterized in that, The calculation process of the self-attention calculation module is as follows: the hidden state vectors of each time period are mapped to query vectors, key vectors and value vectors through three independent linear transformation layers; the scaled dot product of the query vector and the key vector is calculated to obtain the attention weight matrix; the attention weight matrix is ​​multiplied by the value vector to obtain the weighted features; the weighted features are summed along the time period dimension to obtain the global encoding vector of the process; wherein, the self-attention calculation module adopts a multi-head attention form, splits the hidden state vector into multiple subspaces, calculates the attention weights independently in each subspace, and concatenates the weighted outputs of each subspace and maps them back to the original dimension through a linear transformation layer.

5. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 1, characterized in that, The fast surface detection data includes surface roughness distribution map data and surface morphology grayscale image data. Before inputting into the convolutional neural network, Z-score normalization is performed on the surface roughness distribution map data and the surface morphology grayscale image data respectively, and their respective pixel values ​​are mapped to a unified dimensionless distribution and then concatenated along the channel dimension to form a multi-channel input.

6. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 1, characterized in that, In the cross-attention fusion module, the global encoding vector of the process is mapped to a query matrix through a linear transformation layer. The surface observable quality feature vector is mapped to a bond matrix through two independent linear transformation layers. AND-value matrix The formula for calculating scaled dot product attention is: in, For querying the matrix, The key matrix, For value matrices, Indicates transpose. Let be the dimension of the key vector. It is a normalized exponential function.

7. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 1, characterized in that, The blade comprehensive quality risk index The calculation formula is: in, This represents the total number of defect types. For defect type index, For the first Preset weighting coefficients for each defect type For the first The probability value of the existence of each defect type For the first Severity level coding for different defect types; when Less than the preset safety threshold When a qualified assessment conclusion is generated, the blade skips the subsequent non-destructive testing process; when Greater than or equal to At that time, an evaluation conclusion requiring detailed non-destructive testing is generated.

8. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 7, characterized in that, when Greater than or equal to In addition, a set of recommended detection area coordinates and a code for recommended detection method are output. The set of recommended detection area coordinates is generated by extracting the process stage identifiers corresponding to the time periods with the highest attention weight in the self-attention calculation module, and determining the spatial coordinate range of the region with the most drastic fluctuation of process parameters on the blade based on the correspondence between the process stage identifiers and the blade geometric model. The recommended detection method encoding is determined based on the defect type with the highest probability of existence output by the multi-label prediction network.

9. The online evaluation method for manufacturing quality of aero-engine blades based on data analysis according to claim 7, characterized in that, When calculating the comprehensive quality risk index of the blade, a nonlinear mapping is performed on the probability values ​​of each defect type: let the probability warning threshold be... For the first Types of defects, when At that time, with replace Participating in the weighted summation operation, among which Represents an exponential function; when hour, It directly participates in the weighted summation operation.

Images