Performance detection system for precision forging motor shaft and gear set

Abstract

The present application relates to a kind of performance detection system for integrated performance of precision forging motor shaft and gear group, specifically relates to the field of performance detection, realize the high-precision, high-efficiency online full detection and intelligent grading of early weak defect of precision forging motor shaft and gear group, by generating and production line noise environment highly consistent synthetic defect data, effectively overcome the sample scarcity problem, and utilize adaptive denoising and multi-modal feature fusion, in strong noise background Stable extraction defect feature, adopt attention mechanism and multi-task uncertainty evaluation, accurately locate anomaly and output reliable decision, simultaneously through closed-loop evolution mechanism, utilize fuzzy sample driving model continuous adaptive optimization, to significantly improve detection precision, robustness and production line adaptability, finally reach quality zero defect control and production efficiency improvement.

Description

Technical Field

[0001] This invention relates to the field of performance testing, and more specifically, to a comprehensive performance testing system for precision forged motor shafts and gear sets. Background Technology

[0002] With the increasing demands for performance and reliability of core components in the high-end equipment manufacturing industry, precision-forged motor shafts and gear sets, as key components of the power transmission system, directly determine the lifespan and stability of end products due to their consistent quality. In large-scale production scenarios, such as the production line of professional manufacturer Yancheng Mingjia Machinery, achieving 100% rapid full inspection and accurate sorting of such workpieces is a crucial process to ensure product quality and meet customers' zero-defect delivery requirements. However, this scenario faces multiple challenges, including fast production cycles, workpiece surfaces often covered with processing oil stains, and significant background mechanical noise. Traditional offline sampling inspection methods, such as coordinate measuring machine (CMM) or magnetic particle testing, while highly accurate, are inefficient and costly, unable to adapt to the pace of online full inspection. Manual visual inspection is powerless to detect internal and early microscopic defects in workpieces. Therefore, the adoption of non-contact, fast, and efficient online automatic inspection technology has become an urgent need in the industry. Among these, vibration or acoustic signal analysis methods show great potential due to their sensitivity to the internal state of structures and fast response speed, but their effectiveness in actual industrial environments still faces severe challenges.

[0003] Currently, the industry has explored online rapid inspection of workpieces to some extent. Among these methods, vibration or acoustic signal-based detection methods have attracted attention due to their non-contact and rapid response characteristics. Typical existing technical solutions usually involve collecting vibration or sound signals from the workpiece under specific excitation, analyzing the overall energy, peak value, or amplitude of the main frequency components in the spectrum, and comparing them with a preset single threshold to determine whether the workpiece is qualified. However, these methods have significant limitations. First, their detection logic is relatively coarse, only responding to obvious, established defects. They are extremely insensitive to early, weak defects such as microscopic inclusions within the material or microcracks in their initiation stage, leading to a high risk of missed detections. Second, these methods cannot distinguish the type of defect. For example, they cannot differentiate between excessive vibration caused by mass imbalance and abnormal noise caused by pitting or root cracks on the tooth surface, thus failing to support more refined quality grading decisions. For example, distinguishing between qualified products, reworkable products, and scrapped products is challenging. Furthermore, in actual production environments with strong background noise, the characteristic signals of early defects are often submerged, making it difficult for traditional spectrum analysis algorithms to achieve stable and robust feature extraction, leading to an increased false alarm rate in the detection system. The more fundamental challenge lies in the fact that the acoustic and vibration characteristic patterns representing different categories and levels of defects are complex and diverse, and methods that rely on expert experience to manually design features and rules are inexhaustible. On the other hand, data-driven intelligent diagnostic models are constrained by the scarcity of defect samples on the production line, resulting in difficulties in model training and insufficient generalization ability. Therefore, existing online acoustic and vibration detection technologies have significant capability gaps in meeting the requirements for high-precision and high-efficiency identification and classification of early weak defects in precision forged motor shafts and gear sets. There is an urgent need for an innovative comprehensive performance detection system that can deeply integrate advanced signal processing and artificial intelligence analysis and adapt to the complex conditions of industrial sites. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a comprehensive performance testing system for precision forged motor shafts and gear sets. The system utilizes a data synthesis module, a feature extraction module, an intelligent decision-making module, and a model evolution module to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: a data synthesis module, a feature extraction module, an intelligent decision-making module, and a model evolution module connected in sequence; wherein, Data synthesis module: It is used to receive the raw acoustic and vibration signals collected from the production line, which contain a small number of defect samples and a large number of normal samples. It combines the model and material batch information of the current workpiece with the background acoustic features extracted from the environmental noise of the production line to drive a conditional generative adversarial network to generate a synthesized acoustic and vibration signal with specified defect type, defect level and background noise feature matching, thereby outputting an extended training dataset with balanced category and intensity. Feature extraction module: This module receives extended training datasets and real-time detected multi-sensor signals from the workpiece. The multi-sensor signals include at least acoustic vibration signals and optical dimension measurements. The acoustic vibration signals are fed into an adaptive noise suppression unit, where background acoustic features extracted from production line environmental noise are used for initial denoising. Then, a multi-scale residual convolutional network is used to extract hierarchical time-frequency features. Simultaneously, the optical dimension measurements are encoded into auxiliary feature vectors, and a gating mechanism is used to fuse the auxiliary feature vectors with the denoised hierarchical time-frequency features, outputting a high-dimensional fused feature tensor. Intelligent decision-making module: It receives high-dimensional fusion feature tensors. First, it calculates the weight distribution of the high-dimensional fusion feature tensors in different time and frequency dimensions through a region attention unit to focus on abnormal feature regions and obtain a weighted high-dimensional fusion feature tensor. Then, it inputs the weighted high-dimensional fusion feature tensor into a multi-task prediction network head and outputs three prediction results in parallel: defect existence classification, defect type classification, and defect severity level regression, as well as their corresponding cognitive uncertainty estimates. Model Evolution Module: This module receives the prediction results and cognitive uncertainty estimates output by the intelligent decision-making module. It marks samples with cognitive uncertainty estimates higher than a preset threshold as fuzzy samples and stores them in a buffer queue. It periodically triggers a human-machine collaborative annotation process to obtain accurate labels for fuzzy samples. It also uses newly labeled samples with accurate labels obtained through the human-machine collaborative annotation process to incrementally learn and update the parameters of the conditional generative adversarial network in the data synthesis module, as well as the network model parameters in the feature extraction module and the intelligent decision-making module. In a preferred embodiment, the specific process of generating the synthesized acoustic and vibration signals using the driving condition generative adversarial network in the data synthesis module is as follows: The raw acoustic and vibration signals received by the data synthesis module constitute a raw acoustic and vibration signal set, which contains a large number of normal samples and a small number of defective samples; at the same time, it receives model and material batch information and background acoustic features extracted from the production line environmental noise. The defect categories and severity of the defect samples are digitally encoded to form defect condition vectors; background acoustic features are then constructed into background acoustic feature vectors. The generator of the conditional generative adversarial network receives a noise vector, a defect condition vector, and a background acoustic feature vector randomly sampled from a standard normal distribution, and outputs a synthesized time-domain acoustic vibration signal as the synthesized acoustic vibration signal. The training of the conditional generative adversarial network is optimized through a composite loss function, which includes: adversarial loss for the discriminator to judge the authenticity of the input signal, condition matching loss for the discriminator's predicted conditions and the differences between the actual conditions, and multi-resolution spectral consistency loss for constraining the statistical characteristics of the generator's output signal and the corresponding real defect signal at multiple time-frequency resolutions. The multi-resolution spectral consistency loss is achieved by comparing the difference between the logarithmic energy spectrum mean of the generated signal and the corresponding real signal at a set of complex continuous wavelet transform scales distributed in octave bands. The generator’s total loss function is the sum of adversarial loss, conditional matching loss and multi-resolution spectral consistency loss, wherein the conditional matching loss and multi-resolution spectral consistency loss are multiplied by adjustable weighting coefficients before being added together.

[0006] In a preferred embodiment, the complex continuous wavelet transform used in the multi-resolution spectral consistency loss employs a scale parameter selection rule that ensures adjacent scales are in a 2:1 relationship, forming a set of analysis scales distributed in octave bands. The specific calculation process of multi-resolution spectral consistency loss is as follows: Perform a set of complex continuous wavelet transforms on the synthesized acoustic vibration signal output by the generator and the corresponding real defect acoustic vibration signal respectively; take the square of the modulus of the coefficients obtained by the transform at each scale to obtain the energy distribution at each scale; calculate the mean of the energy distribution of the generated signal and the real signal at each scale on the batch data respectively; perform logarithmic operation on the calculated mean of the energy distribution at each scale respectively; finally, calculate the sum of the absolute values ​​of the differences between the generated signal and the real signal after logarithmic operation at all scales. The data synthesis module ultimately outputs an extended training dataset that balances categories and intensities. This extended training dataset consists of the original set of acoustic and vibration signals, as well as synthesized acoustic and vibration signals generated by the generator, which are labeled with corresponding defect categories and severity levels, and background noise features.

[0007] In a preferred embodiment, the specific process of preliminary noise reduction using background acoustic features extracted from production line environmental noise in the feature extraction module is as follows: First, a trainable spectral attention network is constructed, which takes background acoustic features extracted from production line environmental noise as conditional input. This network contains at least two cascaded fully connected layers. The ReLU activation function is applied after the first fully connected layer, and the Sigmoid activation function is applied after the last fully connected layer, ultimately outputting a time-frequency mask. Next, a short-time Fourier transform is performed on the input acoustic signal to obtain its complex spectral representation. Then, the value at each position in the time-frequency mask is subtracted from the numerical value to obtain a noise suppression mask. The noise suppression mask is then multiplied element-wise with the complex spectrum of the acoustic signal to attenuate the frequency components in the complex spectrum corresponding to the noise patterns represented by the background acoustic features. Finally, an inverse short-time Fourier transform is performed on the attenuated complex spectrum after element-wise multiplication to convert it back to a time-domain signal, obtaining the preliminarily denoised acoustic signal.

[0008] In a preferred embodiment, the specific process of extracting hierarchical time-frequency features using a multi-scale residual convolutional network is as follows: The multi-scale residual convolutional network consists of K cascaded feature extraction blocks. For the k-th feature extraction block, its input is the output of the (k-1)-th block, and the input of the zeroth feature extraction block is the pre-denoised acoustic vibration signal. Each feature extraction block contains two parallel convolutional branches. The first branch processes the input feature using a one-dimensional convolution with a dilation rate of 1, and the second branch processes the same input feature using a one-dimensional dilated convolution with a dilation rate of 2 to the power of k. Each convolutional branch sequentially performs one-dimensional convolution, batch normalization, and ReLU activation function operations. The outputs of the two parallel branches within the same feature extraction block are concatenated along the feature channel dimension to form the output feature of that feature extraction block. Finally, the multi-scale residual convolutional network concatenates the outputs of all K feature extraction blocks along the feature channel dimension to form hierarchical time-frequency features.

[0009] In a preferred embodiment, the specific process of fusing the auxiliary feature vector with the denoised hierarchical time-frequency features through a gating mechanism is as follows: First, the optical dimension measurements are encoded into auxiliary feature vectors through a fully connected network. Then, the encoded auxiliary feature vectors are input into a gated generator, which performs a linear transformation on the auxiliary feature vectors and inputs the result of the linear transformation into a Sigmoid activation function to calculate and generate a gated weight vector. The number of scalar elements contained in the gated weight vector is the same as the number of feature channels contained in the hierarchical time-frequency features. Finally, each feature channel of the hierarchical time-frequency features is multiplied numerically with the scalar element with the same index position in the gated weight vector to obtain the weighted result of the corresponding feature channel. The weighted results of all feature channels together constitute a high-dimensional fusion feature tensor.

[0010] In a preferred embodiment, the specific process by which the regional attention unit calculates the weight distribution to focus on the abnormal feature region in the intelligent decision-making module is as follows: First, a lightweight auxiliary binary classifier is constructed. This classifier receives the feature vector obtained by global average pooling of the high-dimensional fused feature tensor in the time dimension and outputs a preliminary probability value indicating that the detected object corresponding to the high-dimensional fused feature tensor has a defect. Next, a Sigmoid activation function is applied to each element value in the high-dimensional fused feature tensor, mapping it to the range of zero to one, obtaining a virtual binary prediction corresponding to that element position. Then, using the preliminary probability value output by the auxiliary binary classifier as the target label and the virtual binary prediction corresponding to each element position as the prediction value for that position, the binary cross-entropy loss value at each element position is calculated. Subsequently, for... For each element position, the calculated binary cross-entropy loss value is negativeed, and then the negative result is multiplied by a preset sharpening coefficient. The result of the multiplication is then subjected to exponential operation with the natural constant e as the base. After that, the results of the exponential operation with the natural constant e at all element positions of the high-dimensional fusion feature tensor are summed, and the result of the exponential operation at each position is divided by the sum to obtain the attention weight corresponding to each element position. All these attention weights together form a weight distribution map with the same dimension as the high-dimensional fusion feature tensor. The weight distribution map is multiplied element-wise with the input high-dimensional fusion feature tensor to obtain the weighted high-dimensional fusion feature tensor.

[0011] In a preferred embodiment, the specific process by which the multi-task prediction network head outputs three prediction results and their corresponding cognitive uncertainty estimates is as follows: The multi-task prediction network head contains a shared low-level feature extraction layer and branches into three independent sub-networks at higher levels, corresponding to the defect existence classification, defect type classification, and defect severity level regression tasks, respectively. The defect existence classification subnetwork outputs a scalar between zero and one, representing the probability that the detected workpiece has a defect corresponding to the weighted high-dimensional fused feature tensor; the defect type classification subnetwork outputs a multi-dimensional vector, the dimension of which is equal to the total number of predefined defect categories, and the value of each element in the multi-dimensional vector represents the probability that the detected workpiece belongs to the corresponding defect category; the defect severity regression subnetwork outputs a scalar greater than or equal to zero, representing the numerical value of the defect severity of the detected workpiece. During the inference phase, the random dropout layer in the network is kept active, and multiple independent forward propagation calculations are performed on the same weighted high-dimensional fusion feature tensor. Each calculation yields a set of prediction results consisting of the probability of defect existence, the probability distribution of defect type, and the numerical value of defect severity level due to the introduction of random dropout. For the defect existence classification task, the cognitive uncertainty is estimated by calculating the average information entropy of the defect existence probability values ​​obtained from multiple predictions. That is, firstly, based on the defect existence probability value obtained from each forward propagation, the information entropy corresponding to that prediction is calculated. This information entropy is equal to the negative defect existence probability value multiplied by the natural logarithm of the defect existence probability value, and the difference between the negative one minus the defect existence probability value multiplied by the natural logarithm of the difference, and the sum of the two. Then, the arithmetic mean of the information entropy calculated from all forward propagations is obtained. For the defect type classification task, the cognitive uncertainty is estimated by calculating the average information entropy of the type probability distribution obtained from multiple predictions. That is, firstly, based on the defect type probability distribution vector obtained from each forward propagation, the information entropy corresponding to the prediction is calculated. This information entropy is equal to the sum of the products of each probability value in the negative defect type probability distribution vector and its natural logarithm. Then, the arithmetic mean of the information entropy calculated from all forward propagations is obtained. For a severity level regression task, the cognitive uncertainty is estimated by calculating the standardized interquartile range (IIR) of the severity levels obtained from multiple predictions. The specific process is as follows: First, the severity levels obtained from multiple predictions are sorted in ascending order; then, the upper and lower quartiles of this sorted sequence are found; next, the difference between the upper and lower quartiles is calculated; then, the sum of the upper and lower quartiles is calculated; finally, the difference is divided by the sum and a positive constant; the quotient is the standardized IIR, which serves as the estimate of the cognitive uncertainty of the regression task. The multi-task prediction network head outputs in parallel the final probability of defect existence, the final probability distribution of defect type, the final regression value of defect severity level, and the corresponding three cognitive uncertainty estimates.

[0012] In a preferred embodiment, the specific process of marking samples with cognitive uncertainty estimates higher than a preset threshold as fuzzy samples and storing them in a buffer queue in the model evolution module is as follows: First, the cognitive uncertainty estimates for defect existence classification, defect type classification, and defect severity level regression output by the intelligent decision-making module are processed using nonlinear transformations. Specifically, the cognitive uncertainty estimates for defect existence classification and defect type classification are processed using a first nonlinear function, which is in the form of: an exponential function of input value multiplied by a negative scaling factor with the natural constant e as the base. The cognitive uncertainty estimate for defect severity level regression is processed using a second nonlinear function, which is in the form of: the arctangent value of constant 2 divided by pi and then multiplied by the input value multiplied by the scaling factor. Next, for the cognitive uncertainty estimates of defect existence classification, defect type classification, and defect severity regression after processing by the first nonlinear function, a first weight coefficient, a second weight coefficient, and a third weight coefficient are assigned, respectively. The first weight coefficient is multiplied by the cognitive uncertainty estimate of defect existence classification after processing by the first nonlinear function to obtain the first product; the second weight coefficient is multiplied by the cognitive uncertainty estimate of defect type classification after processing to obtain the second product; the third weight coefficient is multiplied by the cognitive uncertainty estimate of defect severity regression after processing to obtain the third product; the first, second, and third products are added together to obtain a comprehensive uncertainty score. Then, a dynamic threshold is calculated as follows: multiply the smoothing coefficient by the average comprehensive uncertainty score of the historical buffer queue to obtain the first term; subtract the smoothing coefficient from one to obtain a difference coefficient; calculate the larger value between the average comprehensive uncertainty score of the current buffer queue and a preset minimum threshold; multiply the difference coefficient by the larger value to obtain the second term; finally, add the first term and the second term together, and the sum is the dynamic threshold. Finally, the comprehensive uncertainty score corresponding to the data currently being processed and output by the intelligent decision module is compared with the dynamic threshold. If the comprehensive uncertainty score is greater than the dynamic threshold, this set of data and its corresponding detection object are marked as fuzzy samples, and they, along with the corresponding original acoustic vibration signal, optical size measurement value, background acoustic feature vector and prediction results from the intelligent decision module, are stored in the buffer queue.

[0013] In a preferred embodiment, the specific process of incrementally learning and updating the network model parameters in the data synthesis module, the feature extraction module, and the intelligent decision-making module using newly labeled samples is as follows: First, the newly labeled samples with precise labels obtained through the human-machine collaborative labeling process are processed to generate soft labels for the classification task. The soft label is a weighted combination of the precise label and the original prediction result of the intelligent decision module for the sample, where the weight of the precise label is higher than the weight of the original prediction result. The specific method for generating soft labels is as follows: For the defect type classification task, the defect type in the precise label is converted into a one-hot encoded vector, the probability distribution vector of the defect type output by the intelligent decision module for the sample is multiplied by a preset distillation coefficient, the one-hot encoded vector is multiplied by the difference between the one-hot and distillation coefficients, and the two product results are added together. The sum is the soft label used for training. Subsequently, based on the newly labeled samples and their soft labels, the network model parameters in the feature extraction module and the intelligent decision-making module are updated. The loss function used for the update consists of two parts: the first part is the standard task loss on the newly labeled samples, and the second part is a penalty term for each trainable parameter in the network model. The penalty term is calculated as follows: for each trainable parameter in the network model, a non-negative importance metric representing the importance of the parameter to the learned historical tasks is obtained, and the importance metric, a preset global regularization strength coefficient, and the square of the difference between the current value of the parameter and the old value saved before the start of this incremental learning are multiplied together. The resulting product is the penalty term for the parameter. The penalty terms for all trainable parameters are summed to obtain the second part of the loss function. Meanwhile, the conditional generative adversarial network in the data synthesis module is updated while keeping its discriminator parameters unchanged. The generator is fine-tuned using only real defect samples and their labels from the newly labeled samples. During the fine-tuning process, the optimization objective of the generator is to make the conditions of its generated synthetic acoustic and vibration signals match the defect types and severity level labels in the newly labeled samples, so that it can learn to synthesize such newly confirmed defect patterns.

[0014] The beneficial effects of this invention are: it achieves high-precision and high-efficiency online full inspection and intelligent grading of early weak defects in precision forging motor shafts and gear sets; it effectively overcomes the problem of sample scarcity by generating synthetic defect data that is highly consistent with the noise environment of the production line; it stably extracts defect features under strong noise background by using adaptive denoising and multimodal feature fusion; it accurately locates anomalies and outputs reliable decisions by using attention mechanism and multi-task uncertainty assessment; and it continuously optimizes the model by using fuzzy sample-driven continuous adaptive optimization through closed-loop evolution mechanism, thereby significantly improving detection accuracy, robustness and production line adaptability, and ultimately achieving zero-defect quality control and improved production efficiency. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1-2 The system shown is a comprehensive performance testing system for precision forged motor shafts and gear sets, specifically comprising: a data synthesis module, a feature extraction module, an intelligent decision-making module, and a model evolution module connected in sequence; wherein, Data synthesis module: It is used to receive the raw acoustic and vibration signals collected from the production line, which contain a small number of defect samples and a large number of normal samples. It combines the model and material batch information of the current workpiece with the background acoustic features extracted from the environmental noise of the production line to drive a conditional generative adversarial network to generate a synthesized acoustic and vibration signal with specified defect type, defect level and background noise feature matching, thereby outputting an extended training dataset with balanced category and intensity. Feature extraction module: This module receives extended training datasets and real-time detected multi-sensor signals from the workpiece. The multi-sensor signals include at least acoustic vibration signals and optical dimension measurements. The acoustic vibration signals are fed into an adaptive noise suppression unit, where background acoustic features extracted from production line environmental noise are used for initial denoising. Then, a multi-scale residual convolutional network is used to extract hierarchical time-frequency features. Simultaneously, the optical dimension measurements are encoded into auxiliary feature vectors, and a gating mechanism is used to fuse the auxiliary feature vectors with the denoised hierarchical time-frequency features, outputting a high-dimensional fused feature tensor. Intelligent decision-making module: It receives high-dimensional fusion feature tensors. First, it calculates the weight distribution of the high-dimensional fusion feature tensors in different time and frequency dimensions through a region attention unit to focus on abnormal feature regions and obtain a weighted high-dimensional fusion feature tensor. Then, it inputs the weighted high-dimensional fusion feature tensor into a multi-task prediction network head and outputs three prediction results in parallel: defect existence classification, defect type classification, and defect severity level regression, as well as their corresponding cognitive uncertainty estimates. Model Evolution Module: This module receives the prediction results and cognitive uncertainty estimates output by the intelligent decision-making module. It marks samples with cognitive uncertainty estimates higher than a preset threshold as fuzzy samples and stores them in a buffer queue. It periodically triggers a human-machine collaborative annotation process to obtain accurate labels for fuzzy samples. It then uses the newly labeled samples with accurate labels obtained through the human-machine collaborative annotation process to incrementally learn and update the parameters of the conditional generative adversarial network in the data synthesis module, as well as the network model parameters in the feature extraction module and the intelligent decision-making module.

[0020] In this embodiment, it is necessary to specifically explain the process by which the driving condition generative adversarial network generates the synthesized acoustic and vibration signals in the data synthesis module: The data synthesis module receives the raw acoustic and vibration signals collected from the production line, forming a raw acoustic and vibration signal set. This raw acoustic and vibration signal set contains a large number of normal samples representing a defect-free state and a small number of defect samples containing various types of defects. At the same time, it receives the model and material batch information representing the workpiece specifications, as well as the background acoustic features extracted from the production line environmental noise. Based on the defect category and severity corresponding to the defect samples in the original acoustic and vibration signal set, digital encoding is performed to form a unified defect condition vector; background acoustic features extracted from the production line environmental noise are constructed into a background acoustic feature vector. The generator of the conditional generative adversarial network receives three inputs: a noise vector randomly sampled from a standard normal distribution, a defect condition vector, and a background acoustic feature vector; based on these three inputs, the generator outputs a synthesized time-domain acoustic vibration signal as the synthesized acoustic vibration signal. The training process of a conditional generative adversarial network (GAN) optimizes the generator and discriminator through a composite loss function. This composite loss function consists of three parts: the first part is the adversarial loss for the discriminator to determine the authenticity of the received acoustic signal; the second part is the conditional matching loss, which measures the difference between the discriminator's predicted defect condition vector and background acoustic feature vector corresponding to the received signal and the actual defect condition vector and background acoustic feature vector corresponding to the signal; the third part is the multi-resolution spectral consistency loss, which constrains the synthesized acoustic signal output by the generator to match the defect condition vectors in the original acoustic signal set that have the same characteristics as the synthesized acoustic signal. The statistical characteristics of the acoustic vibration signals corresponding to the defect samples in the background acoustic feature vector remain matched at multiple different time-frequency resolutions. The multi-resolution spectral consistency loss is achieved by comparing the difference between the logarithmic energy spectrum mean of the generated signal and the corresponding real signal at a set of complex continuous wavelet transform scales distributed in octave bands. This design forces the generator to imitate the spectral structure of the real defect signal at multiple time-frequency granularities. In particular, it can effectively constrain the energy distribution of subtle resonance features limited to specific frequency bands caused by weak defects (such as microcracks), thereby generating synthetic samples with clear physical meaning and more deceptive characteristics, which significantly improves the data augmentation quality for early defects in the context of strong noise. During generator training, the overall optimization objective is a total loss function, which is the sum of adversarial loss, conditional matching loss, and multi-resolution spectral consistency loss. The conditional matching loss and multi-resolution spectral consistency loss are each multiplied by an adjustable weighting coefficient before summing. For example, the coefficient for conditional matching loss can be set to 1.0, and the coefficient for multi-resolution spectral consistency loss to 0.1. Adjusting the latter controls the strictness of the matching between the generated signal and the real signal in terms of time-spectral statistical characteristics. In the early stages of model training, the weight of multi-resolution spectral consistency loss can be appropriately reduced to stabilize adversarial training; in the later stages of training, its weight can be gradually increased to refine the spectral characteristics of the generated signal. The specific calculation process of multi-resolution spectral consistency loss is as follows: A series of complex continuous wavelet transforms are performed on the synthesized acoustic vibration signal output by the generator and the acoustic vibration signals corresponding to the defect samples in the original acoustic vibration signal set that have the same defect condition vector and background acoustic feature vector as the synthesized acoustic vibration signal. Each transform corresponds to a specified scale parameter, which is distributed according to an octave band pattern. The modulus of the coefficients obtained from the transform at each scale is squared to obtain the energy distribution at each scale. Then, the acoustic vibration signals corresponding to the defect samples in the original acoustic vibration signal set that have the same defect condition vector and background acoustic feature vector as the synthesized acoustic vibration signal are calculated on the batch data. The mean of energy distribution at each scale is calculated. Then, logarithmic operations are performed on the mean energy distribution of the synthesized acoustic signal output by the generator and the acoustic signals corresponding to the defect samples in the original acoustic signal set that have the same defect condition vector and background acoustic feature vector as the synthesized acoustic signal at each scale. Before the logarithmic operation, a very small positive coefficient is added to each mean. Finally, the results of the above logarithmic operation on all scales are successively subtracted from the synthesized acoustic signal output by the generator and the acoustic signals corresponding to the defect samples in the original acoustic signal set that have the same defect condition vector and background acoustic feature vector as the synthesized acoustic signal. The sum of the absolute values ​​of these differences at all scales is then calculated. In the multi-resolution spectral consistency loss, the complex continuous wavelet transform used for multi-resolution analysis of signals selects its scale parameter according to the rule that the scales are twice the size of each other, i.e., the scale parameter of the next scale is twice the scale parameter of the previous scale, thus forming a set of discrete analysis scales distributed in octave bands. When calculating the difference in average energy spectrum, the logarithmic operation used can simulate the characteristics of the human auditory perception system to respond approximately logarithmically to sound intensity, while amplifying the sensitivity to subtle differences in energy in weak frequency bands; after completing the logarithmic operation, the absolute value of the difference between the logarithmic average energy spectra of the two signals is calculated. The generator's total loss function is a weighted sum of adversarial loss, conditional matching loss, and multi-resolution spectral consistency loss. The background acoustic feature vector serves as one of the generator's conditional inputs, enabling the generator to learn to fuse specified environmental noise patterns into the synthesized acoustic and vibration signals. By using the background acoustic feature vector extracted in real time from the production line as a conditional input, the generator can embed the statistical characteristics of specific noises, such as periodic machine tool impacts and air turbulence, into the synthesized signal. This makes the samples in the extended training dataset naturally have a noise background that is highly consistent with the final application scenario (online production line inspection). Thus, when training the diagnostic model, no additional denoising preprocessing or domain adaptation is required, and the model can learn to identify defect features in real noise, greatly improving the system's practicality and robustness. The synthesized acoustic vibration signal output by the generator carries precise defect category and severity labels from the input defect condition vector, as well as background noise feature labels from the input background acoustic feature vector. The data synthesis module ultimately outputs an extended training dataset that balances categories and intensities. This extended training dataset consists of the original set of acoustic and vibration signals, as well as synthesized acoustic and vibration signals generated by the generator, which are labeled with corresponding defect categories and severity levels, and background noise features.

[0021] In this embodiment, it is specifically necessary to explain the preliminary noise reduction process in the feature extraction module, which utilizes the background acoustic features extracted from the production line environmental noise: First, a trainable spectral attention network is constructed, which takes background acoustic features extracted from production line environmental noise as conditional input. This network contains at least two cascaded fully connected layers. A ReLU activation function is applied after the first fully connected layer, and a Sigmoid activation function is applied after the last fully connected layer. The final output is a time-frequency mask with values ​​ranging from zero to one. The dimension of this mask corresponds to the dimension of the complex spectrum obtained after the short-time Fourier transform of the acoustic signal. The Sigmoid activation function ensures that each element of the output time-frequency mask is in the (0,1) interval. The closer the value is to 1, the greater the probability that the corresponding time-frequency point is dominated by background noise, thus resulting in stronger attenuation at that point when calculating the noise suppression mask (1-M) subsequently. Next, a short-time Fourier transform is performed on the input acoustic signal to obtain its complex spectrum representation. The short-time Fourier transform can use a Hanning window with a window length of 256 sampling points and an overlap rate of 50%. The noise suppression mask is obtained by subtracting the value at each position in the time-frequency mask from the value of 1. The noise suppression mask is then multiplied element-wise with the complex spectrum of the acoustic signal to attenuate the frequency components in the complex spectrum corresponding to the noise patterns represented by the background acoustic features. Finally, the complex spectrum attenuated by the element-wise multiplication operation is subjected to a short-time Fourier inverse transform to convert it back to the time domain signal, resulting in the pre-denoised acoustic signal. This process enables the noise suppression to adaptively adjust to the specified noise patterns represented by the background acoustic features, rather than using a fixed threshold. This allows for better preservation of transient impact signal components that may represent defects while suppressing steady-state and quasi-steady-state background noise. By reusing the same background acoustic feature vector from the data synthesis module, this denoising process shares noise priors with the data generation process, ensuring the consistency between the synthesized training data and the real detection signal on the noise background. This eliminates the need for additional domain adaptation in subsequent models, significantly improving the generalization ability from simulation to real-world scenarios. The specific process of extracting hierarchical time-frequency features using a multi-scale residual convolutional network is as follows: The multi-scale residual convolutional network consists of K cascaded feature extraction blocks, where K is an integer greater than one. Typical values ​​for K range from 4 to 6, for example, K=5, to construct hierarchical features with sufficient receptive field diversity. For the k-th feature extraction block, k ranges from 1 to K, and its input is the output of the (k-1)-th block. The input of the zeroth feature extraction block is the pre-denoised acoustic vibration signal. Each feature extraction block contains two parallel convolutional branches. The first branch processes the input features using a one-dimensional convolution with a dilation rate of one. The "dilation rate of 2 to the power of k" design ensures that the receptive field of the second branch grows exponentially with the block index k (e.g., 2 for k=1, 4 for k=2, and 8 for k=3), thus enabling simultaneous capture of features caused by local pitting erosion. The instantaneous high-frequency vibrations and low-frequency periodic fluctuations caused by shaft bending, etc., form feature representations covering different time scales. The second branch uses a one-dimensional dilated convolution with an inflation rate of 2 to the power of k to process the same input feature. Each convolutional branch sequentially performs one-dimensional convolution, batch normalization, and ReLU activation function operations. The momentum parameter used in the batch normalization operation can be set to 0.9 to stably estimate the mean and variance of the batch data during training, accelerate network convergence, and improve generalization performance. The outputs of the two parallel branches within the same feature extraction block are concatenated along the feature channel dimension to form the output feature of that feature extraction block. The multi-scale residual convolutional network finally concatenates the outputs of all K feature extraction blocks along the feature channel dimension to form hierarchical time-frequency features. The specific process of fusing the auxiliary feature vector with the denoised hierarchical time-frequency features through a gating mechanism is as follows: First, optical dimension measurements are encoded into auxiliary feature vectors using a fully connected network. Then, these auxiliary feature vectors are input into a gated generator. This generator performs a linear transformation on the auxiliary feature vectors and inputs the result into a Sigmoid activation function to generate a gated weight vector. This gated weight vector contains the same number of scalar elements as the hierarchical time-frequency features, and the value of each scalar element is mapped to a value between zero and one by the Sigmoid activation function. The dimension of the weight matrix of the linear transformation is [auxiliary feature vector dimension, number of hierarchical time-frequency feature channels]. The Sigmoid function constrains each weight to the (0,1) interval. Physically, this means that when optical dimension information (such as roundness error) suggests a possible specific type of imbalance, the weights of the feature channels corresponding to the frequency of that imbalance in the gated weight vector will approach 1, thus enhancing the response of that channel. Conversely, for feature channels unrelated to the current size anomaly, their weights will be suppressed to near 0, achieving dynamic feature selection and enhancement based on physical priors. Finally, each feature channel of the hierarchical time-frequency features is multiplied numerically with the scalar element with the same index position in the gated weight vector to obtain the weighted result of the corresponding feature channel. The weighted results of all feature channels together constitute a high-dimensional fusion feature tensor.

[0022] In this embodiment, it is necessary to specifically explain the process by which the regional attention unit in the intelligent decision-making module calculates the weight distribution to focus on the abnormal feature region: First, a lightweight auxiliary binary classifier is constructed. This auxiliary binary classifier receives the feature vector obtained by global average pooling of the high-dimensional fused feature tensor in the time dimension and outputs a preliminary probability value indicating that the detected object corresponding to the high-dimensional fused feature tensor has a defect. The auxiliary binary classifier can be composed of two fully connected layers, using the ReLU activation function in the middle and the Sigmoid activation function in the output layer. Next, the Sigmoid activation function is applied to each element value in the high-dimensional fused feature tensor, mapping it to the range of zero to one, to obtain the virtual binary prediction corresponding to that element position. Then, using the preliminary probability value output by the auxiliary binary classifier as the target label and the virtual binary prediction corresponding to each element position as the prediction value for that position, the binary cross-entropy loss value at each element position is calculated. Subsequently, for each element position, the calculated binary cross-entropy loss value is taken as... The negative number is then multiplied by a preset sharpening coefficient greater than zero. The typical value of the sharpening coefficient is between 0.5 and 10, for example, it can be set to 5. This controls the concentration of attention weights; the larger the coefficient, the more concentrated the weights are on a few key feature positions. The result of the multiplication is then subjected to an exponential operation with the natural constant e as the base. After this, the results of the exponential operation with the natural constant e as the base are summed at all positions of the high-dimensional fusion feature tensor, and the result of the exponential operation at each position is divided by the sum, so that the sum of the quotients at all positions is one. This gives the attention weight corresponding to each element position, with values ​​between zero and one. All these attention weights together form a weight distribution map with the same dimension as the high-dimensional fusion feature tensor. The weight distribution map is then multiplied element-wise with the input high-dimensional fusion feature tensor to obtain the weighted high-dimensional fusion feature tensor. This process forces attention weights to be correlated with the consistency of the overall initial defect judgment, enabling the model to automatically focus on the feature regions most relevant to the defect decision logic, rather than simply responding to features with large amplitudes. This mechanism forces the model to learn to identify key regions that are consistent with the overall decision logic in their local feature responses by aligning attention weights with an initial, global defect judgment probability. For example, a sudden energy surge in a specific frequency band that is highly correlated with the overall "crack exists" judgment will be given high weight, while a broadband noise with a large amplitude but unrelated to the defect type will be suppressed. This achieves feature focusing based on decision consistency, significantly improving the detection rate and recognition accuracy of weak and localized defect patterns. The specific process of the multi-task prediction network head outputting three prediction results and their corresponding cognitive uncertainty estimates is as follows: The multi-task prediction network head contains a shared low-level feature extraction layer and branches into three independent sub-networks at higher levels, corresponding to the defect existence classification, defect type classification and defect severity level regression tasks, respectively. The shared low-level feature extraction layer may contain one or two fully connected layers, which are used to further compress and abstract the weighted high-dimensional fusion feature tensor. The defect existence classification subnetwork outputs a scalar between zero and one, representing the probability that the detected workpiece has a defect corresponding to the weighted high-dimensional fused feature tensor; the defect type classification subnetwork outputs a multi-dimensional vector, the dimension of which is equal to the total number of predefined defect categories, and the value of each element in the multi-dimensional vector represents the probability that the detected workpiece belongs to the corresponding category of defect, and the sum of all elements is one; the defect severity regression subnetwork outputs a scalar greater than or equal to zero, representing the numerical value of the defect severity of the detected workpiece. During the inference phase, the random dropout layer in the network is kept active, and multiple independent forward propagation calculations are performed on the same weighted high-dimensional fusion feature tensor. Each calculation yields a set of prediction results consisting of the probability of defect existence, the probability distribution of defect type, and the numerical value of defect severity level due to the introduction of random dropout. The number of multiple independent forward propagations is a preset integer greater than one, which is usually between ten and fifty, for example, set to twenty times, in order to achieve a balance between computational cost and the stability of uncertainty estimation. For the defect existence classification task, the cognitive uncertainty is estimated by calculating the average information entropy of the defect existence probability values ​​obtained from multiple predictions. First, based on the defect existence probability value obtained from each forward propagation, the information entropy corresponding to that prediction is calculated. This information entropy is equal to the negative defect existence probability value multiplied by the natural logarithm of the defect existence probability value, and the difference between the negative one minus the defect existence probability value multiplied by the natural logarithm of the difference, and the sum of the two. Then, the arithmetic mean of the information entropy calculated from all forward propagations is taken. The larger the information entropy value, the more hesitant the model is in judging whether a defect exists, and the higher the cognitive uncertainty estimate is. For the defect type classification task, the cognitive uncertainty is estimated by calculating the average information entropy of the type probability distribution obtained from multiple predictions. That is, firstly, based on the defect type probability distribution vector obtained from each forward propagation, the information entropy corresponding to the prediction is calculated. This information entropy is equal to the sum of the products of each probability value in the negative defect type probability distribution vector and its natural logarithm. Then, the arithmetic mean of the information entropy calculated from all forward propagations is obtained. For a severity level regression task, the cognitive uncertainty is estimated by calculating the standardized interquartile range (IIR) of the severity levels obtained from multiple predictions. The specific process is as follows: First, the severity levels obtained from multiple predictions are sorted in ascending order. Then, the upper and lower quartiles of this sorted sequence are identified, where the upper quartile is the value at the 75th percentile and the lower quartile is the value at the 25th percentile. Next, the difference between the upper and lower quartiles is calculated. Then, the sum of the upper and lower quartiles is calculated. Finally, the difference is divided by the sum of the two quartiles and a positive constant. The positive constant is used to ensure that the denominator is not zero; it is a very small number, such as 1 multiplied by 10 to the power of -8 (1e-8). The standardized interquartile range is a dimensionless value between zero and one. The closer its value is to one, the higher the dispersion of the multiple predictions relative to its central value, and the more uncertain the model's prediction of the regression value. The quotient is the standardized interquartile range, which serves as an estimate of the cognitive uncertainty of the regression task. The multi-task prediction network head outputs in parallel the final probability of defect existence, the final probability distribution of defect type, the final regression value of defect severity level, and the corresponding three cognitive uncertainty estimates.

[0023] In this embodiment, it is specifically necessary to explain the process in the model evolution module of marking samples with cognitive uncertainty estimates higher than a preset threshold as fuzzy samples and storing them in a buffer queue as follows: First, the cognitive uncertainty estimates for defect existence classification, defect type classification, and defect severity level regression output by the intelligent decision-making module are subjected to nonlinear transformation processing. A scaling factor is used to control the sensitivity and saturation point of the nonlinear function; for example, a scaling factor of three can be used for classification uncertainty, and a scaling factor of five can be used for regression uncertainty. Specifically, the cognitive uncertainty estimates for defect existence classification and defect type classification are processed using a first nonlinear function, which takes the form of: one minus a negative scaling factor with the natural constant e as its base. The input value is multiplied by an exponential function, where the scaling factor is a preset constant greater than zero. This function ensures that the output value grows almost linearly when the input value is small and gradually approaches one when the input value is large, effectively distinguishing between low, medium, and high levels of uncertainty. For the cognitive uncertainty estimation of defect severity regression, a second nonlinear function is used. The form of the second nonlinear function is: constant two divided by pi, then multiplied by the arctangent value of the input value multiplied by the scaling factor, where the scaling factor is a preset constant greater than zero. Both the first and second nonlinear functions map their input values ​​to a numerical range greater than or equal to zero and less than one. Next, for the cognitive uncertainty estimation of defect existence classification after processing with the first nonlinear function, the cognitive uncertainty estimation of defect type classification after processing with the first nonlinear function, and the cognitive uncertainty estimation of defect severity regression after processing with the second nonlinear function, first, second, and third weight coefficients are assigned respectively. All three weight coefficients are positive numbers, and their sum is one. The weight coefficients can be adjusted according to the application stage. For example, in the initial stage of system deployment, the first weight coefficient (corresponding to defect existence) can be set to 0.5, and the second weight coefficient (corresponding to defect type) to... The first weighting coefficient is 0.3, and the third weighting coefficient (corresponding to the severity level) is 0.2 to prioritize defect detection. After the system stabilizes, the coefficients can be adjusted to 0.3, 0.5, and 0.2 to optimize defect identification accuracy. The first weighting coefficient is multiplied by the cognitive uncertainty estimate of the processed defect existence classification to obtain the first product. The second weighting coefficient is multiplied by the cognitive uncertainty estimate of the processed defect type classification to obtain the second product. The third weighting coefficient is multiplied by the cognitive uncertainty estimate of the processed defect severity level regression to obtain the third product. The first, second, and third products are added together to obtain a comprehensive uncertainty score between zero and one. Then, a dynamic threshold is calculated as follows: multiply a smoothing coefficient between zero and one by the average comprehensive uncertainty score of the historical buffer queue to obtain the first term; subtract the smoothing coefficient from one to obtain a difference coefficient; calculate the larger value between the current average comprehensive uncertainty score of the buffer queue and a preset minimum threshold, which can be set to 0.2 to prevent the collection of low-value samples when the data is too simple; the smoothing coefficient can be set to 0.7 to give a higher weight to the historical baseline and make the threshold change more smoothly; multiply the difference coefficient by the larger value to obtain the second term; finally, add the first term and the second term together, and the sum is the dynamic threshold. Finally, the comprehensive uncertainty score corresponding to the data currently being processed and output by the intelligent decision-making module is compared with the dynamic threshold. If the comprehensive uncertainty score is greater than the dynamic threshold, this set of data and its corresponding detection object are marked as fuzzy samples, and they, along with the corresponding original acoustic and vibration signals, optical size measurements, background acoustic feature vectors, and prediction results from the intelligent decision-making module, are stored in a buffer queue. This process, through multi-dimensional fusion and adaptive threshold setting, achieves precise screening of samples that are most perplexing to the model and have the highest evolutionary value. This mechanism, by fusing the uncertainty of the model in three key decision dimensions and using dynamic thresholds for filtering, can adapt to changes in the distribution of production line data. For example, when new defects or noise patterns suddenly appear, the model uncertainty generally increases, and the dynamic threshold will be adjusted accordingly to avoid collecting too many simple samples. When the model's judgment on a certain type of known defect is stable, the dynamic threshold can effectively screen out those truly difficult samples that are at the decision boundary and have fuzzy features, providing the most valuable targets for human-machine collaborative annotation, thereby significantly improving the efficiency of system evolution. The specific process of incrementally learning and updating the network model parameters in the data synthesis module, the feature extraction module, and the intelligent decision-making module using newly labeled samples is as follows: First, the newly labeled samples with precise labels obtained through the human-machine collaborative annotation process are processed to generate soft labels for the classification task. The soft label is a weighted combination of the precise label and the original prediction result of the intelligent decision module for the sample, where the weight of the precise label is higher than the weight of the original prediction result. The specific method for generating soft labels is as follows: For the defect type classification task, the defect type in the precise label is converted into a one-hot encoded vector. The probability distribution vector of the defect type output by the intelligent decision module for the sample is multiplied by a preset distillation coefficient with a value between zero and one. The distillation coefficient is usually set to a small value, such as 0.1, which means that in the soft label, the expert precise label is dominant (weight 0.9), while the original prediction of the model is used as auxiliary information (weight 0.1), so as to achieve knowledge distillation and make the model update smoother. The one-hot encoded vector is multiplied by the difference between one and the distillation coefficient, and the two product results are added together. The sum is the soft label used for training. Subsequently, based on the newly labeled samples and their soft labels, the network model parameters in the feature extraction module and the intelligent decision-making module are updated. The loss function used for the update consists of two parts: the first part is the standard task loss on the newly labeled samples, and the second part is a penalty term for each trainable parameter in the network model. The penalty term is calculated as follows: for each trainable parameter in the network model, a non-negative importance metric is obtained, representing the parameter's importance to the learned historical tasks. The importance metric can be estimated by calculating the diagonal elements of the Fisher information matrix corresponding to the parameter. The larger the value, the steeper the loss function surface of the parameter on the historical tasks, and the greater the potential impact of changing the parameter on historical performance. This importance metric and the... The penalty term for a parameter is calculated by multiplying the three factors: a preset global regularization strength coefficient, the square of the difference between the current value of the parameter and the old value saved before the start of this incremental learning, and the sum of the penalty terms for all trainable parameters. This yields the second part of the loss function. The global regularization strength coefficient can be used to control the balance between the retention of new and old knowledge; for example, it can be set to 1000. This loss function design ensures that when learning from new samples, the changes of parameters that are crucial to the historical task are strongly constrained, while parameters that are not important to the historical task can be updated relatively freely to adapt to new knowledge, thus effectively preventing catastrophic forgetting. By minimizing this loss function, the model adapts to new samples while strictly constraining the changes of parameters that are important to the historical task to prevent catastrophic forgetting. Simultaneously, the conditional generative adversarial network in the data synthesis module is updated while keeping its discriminator parameters unchanged. The generator is fine-tuned using only real defect samples and their labels from the newly labeled samples. During the fine-tuning process, the generator's optimization objective includes matching the conditions of its generated synthetic acoustic and vibration signals with the defect types and severity level labels in the newly labeled samples, enabling it to learn to synthesize such newly confirmed defect patterns. This update strategy aims to expand the defect pattern library of the data synthesis module. By using only a small number of newly added real defect samples to fine-tune the generator through a few iterations, it can learn to synthesize new defect patterns that are distributed in the same way as the newly added samples, based on its original generation capabilities. This continuously enriches and expands the diversity and realism of the training dataset, providing a higher quality data foundation for the continuous evolution of the entire system.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A comprehensive performance testing system for precision-forged motor shafts and gear sets, characterized in that, Specifically, it includes: The data synthesis module, feature extraction module, intelligent decision-making module, and model evolution module are connected sequentially; among them, Data synthesis module: It is used to receive the raw acoustic and vibration signals collected from the production line, which contain a small number of defect samples and a large number of normal samples. It combines the model and material batch information of the current workpiece with the background acoustic features extracted from the environmental noise of the production line to drive a conditional generative adversarial network to generate a synthesized acoustic and vibration signal with specified defect type, defect level and background noise feature matching, thereby outputting an extended training dataset with balanced category and intensity. Feature extraction module: This module receives extended training datasets and real-time detected multi-sensor signals from the workpiece. The multi-sensor signals include at least acoustic vibration signals and optical dimension measurements. The acoustic vibration signals are fed into an adaptive noise suppression unit, where background acoustic features extracted from production line environmental noise are used for initial denoising. Then, a multi-scale residual convolutional network is used to extract hierarchical time-frequency features. Simultaneously, the optical dimension measurements are encoded into auxiliary feature vectors, and a gating mechanism is used to fuse the auxiliary feature vectors with the denoised hierarchical time-frequency features, outputting a high-dimensional fused feature tensor. Intelligent decision-making module: It receives high-dimensional fusion feature tensors. First, it calculates the weight distribution of the high-dimensional fusion feature tensors in different time and frequency dimensions through a region attention unit to focus on abnormal feature regions and obtain a weighted high-dimensional fusion feature tensor. Then, it inputs the weighted high-dimensional fusion feature tensor into a multi-task prediction network head and outputs three prediction results in parallel: defect existence classification, defect type classification, and defect severity level regression, as well as their corresponding cognitive uncertainty estimates. Model Evolution Module: This module receives the prediction results and cognitive uncertainty estimates output by the intelligent decision-making module. It marks samples with cognitive uncertainty estimates higher than a preset threshold as fuzzy samples and stores them in a buffer queue. It periodically triggers a human-machine collaborative annotation process to obtain accurate labels for fuzzy samples. It then uses the newly labeled samples with accurate labels obtained through the human-machine collaborative annotation process to incrementally learn and update the parameters of the conditional generative adversarial network in the data synthesis module, as well as the network model parameters in the feature extraction module and the intelligent decision-making module.

2. The comprehensive performance testing system for precision forged motor shafts and gear sets according to claim 1, characterized in that: In the data synthesis module, the specific process of generating the synthesized acoustic and vibration signals using the driving conditional generative adversarial network is as follows: The raw acoustic and vibration signals received by the data synthesis module constitute a raw acoustic and vibration signal set, which contains a large number of normal samples and a small number of defective samples; at the same time, it receives model and material batch information and background acoustic features extracted from the production line environmental noise. The defect category and severity of the defect samples are digitally encoded to form a defect condition vector; Construct background acoustic features into a background acoustic feature vector; The generator of the conditional generative adversarial network receives a noise vector, a defect condition vector, and a background acoustic feature vector randomly sampled from a standard normal distribution, and outputs a synthesized time-domain acoustic vibration signal as the synthesized acoustic vibration signal. The training of the conditional generative adversarial network is optimized through a composite loss function, which includes: adversarial loss for the discriminator to judge the authenticity of the input signal, condition matching loss for the discriminator's predicted conditions and the differences between the actual conditions, and multi-resolution spectral consistency loss for constraining the statistical characteristics of the generator's output signal and the corresponding real defect signal at multiple time-frequency resolutions. The multi-resolution spectral consistency loss is achieved by comparing the difference between the logarithmic energy spectrum mean of the generated signal and the corresponding real signal at a set of complex continuous wavelet transform scales distributed in octave bands. The generator’s total loss function is the sum of adversarial loss, conditional matching loss and multi-resolution spectral consistency loss, wherein the conditional matching loss and multi-resolution spectral consistency loss are multiplied by adjustable weighting coefficients before being added together.

3. The comprehensive performance testing system for precision-forged motor shafts and gear sets according to claim 2, characterized in that: The complex continuous wavelet transform used in the multi-resolution spectral consistency loss employs a scale parameter selection rule that ensures adjacent scales are in a 2:1 relationship, forming a set of analysis scales distributed in octave bands. The specific calculation process of multi-resolution spectral consistency loss is as follows: Perform a set of complex continuous wavelet transforms on the synthesized acoustic vibration signal output by the generator and the corresponding real defect acoustic vibration signal respectively; take the square of the modulus of the coefficients obtained by the transform at each scale to obtain the energy distribution at each scale; calculate the mean of the energy distribution of the generated signal and the real signal at each scale on the batch data respectively; perform logarithmic operation on the calculated mean of the energy distribution at each scale respectively; finally, calculate the sum of the absolute values ​​of the differences between the generated signal and the real signal after logarithmic operation at all scales. The data synthesis module ultimately outputs an extended training dataset that balances categories and intensities. This extended training dataset consists of the original set of acoustic and vibration signals, as well as synthesized acoustic and vibration signals generated by the generator, which are labeled with corresponding defect categories and severity levels, and background noise features.

4. The comprehensive performance testing system for precision-forged motor shafts and gear sets according to claim 3, characterized in that: The specific process of preliminary noise reduction using background acoustic features extracted from production line environmental noise in the feature extraction module is as follows: First, a trainable spectral attention network is constructed, which takes background acoustic features extracted from production line environmental noise as conditional input. This spectral attention network contains at least two cascaded fully connected layers. After the first fully connected layer, a ReLU activation function is applied, and after the last fully connected layer, a Sigmoid activation function is applied, ultimately outputting a time-frequency mask. Next, a short-time Fourier transform is performed on the input acoustic signal to obtain its complex spectral representation. Then, the value at each position in the time-frequency mask is subtracted from the value to obtain the noise suppression mask. The noise suppression mask is multiplied element-wise with the complex spectrum of the acoustic vibration signal to attenuate the frequency components in the complex spectrum that correspond to the noise patterns characterized by the background acoustic features. Finally, a short-time inverse Fourier transform is performed on the complex spectrum after the element-wise multiplication operation to convert it back to the time domain signal, thus obtaining the acoustic vibration signal after preliminary denoising.

5. The comprehensive performance testing system for precision forged motor shafts and gear sets according to claim 4, characterized in that: The specific process by which the multi-scale residual convolutional network extracts hierarchical time-frequency features is as follows: The multi-scale residual convolutional network consists of K cascaded feature extraction blocks. For the k-th feature extraction block, its input is the output of the (k-1)-th block, and the input of the zeroth feature extraction block is the pre-denoised acoustic vibration signal. Each feature extraction block contains two parallel convolutional branches. The first branch processes the input feature using a one-dimensional convolution with a dilation rate of 1, and the second branch processes the same input feature using a one-dimensional dilated convolution with a dilation rate of 2 to the power of k. Each convolutional branch sequentially performs one-dimensional convolution, batch normalization, and ReLU activation function operations. The outputs of the two parallel branches within the same feature extraction block are concatenated along the feature channel dimension to form the output feature of that feature extraction block. Finally, the multi-scale residual convolutional network concatenates the outputs of all K feature extraction blocks along the feature channel dimension to form hierarchical time-frequency features.

6. The comprehensive performance testing system for precision forged motor shafts and gear sets according to claim 5, characterized in that: The specific process of fusing the auxiliary feature vector with the denoised hierarchical time-frequency features through a gating mechanism is as follows: First, the optical dimension measurements are encoded into auxiliary feature vectors through a fully connected network. Then, the encoded auxiliary feature vectors are input into a gated generator, which performs a linear transformation on the auxiliary feature vectors and inputs the result of the linear transformation into a Sigmoid activation function to calculate and generate a gated weight vector. The number of scalar elements contained in the gated weight vector is the same as the number of feature channels contained in the hierarchical time-frequency features. Finally, each feature channel of the hierarchical time-frequency features is multiplied numerically with the scalar element with the same index position in the gated weight vector to obtain the weighted result of the corresponding feature channel. The weighted results of all feature channels together constitute a high-dimensional fusion feature tensor.

7. The comprehensive performance testing system for precision forged motor shafts and gear sets according to claim 6, characterized in that: In the intelligent decision-making module, the specific process by which the regional attention unit calculates the weight distribution to focus on abnormal feature regions is as follows: First, a lightweight auxiliary binary classifier is constructed. This auxiliary binary classifier receives the feature vector obtained by global average pooling of the high-dimensional fusion feature tensor in the time dimension, and outputs a preliminary probability value indicating that the detected object corresponding to the high-dimensional fusion feature tensor has a defect. Then, a sigmoid activation function is applied to each element value in the high-dimensional fusion feature tensor to map it to the range of zero to one, thereby obtaining the virtual binary prediction corresponding to the element position. Then, using the initial probability value output by the auxiliary binary classifier as the target label, and the virtual binary prediction corresponding to each element position as the predicted value for that position, the binary cross-entropy loss value at each element position is calculated. Subsequently, for each element position, the calculated binary cross-entropy loss value is negativeized, and the negative result is multiplied by a preset sharpening coefficient. The result of the multiplication is then subjected to exponential operation with the natural constant e as the base. After this, the results obtained by the exponential operation with the natural constant e as the base for all element positions of the high-dimensional fusion feature tensor are summed, and the result of the exponential operation of each position is divided by the sum to obtain the attention weight corresponding to each element position. All these attention weights together constitute a weight distribution map with the same dimension as the high-dimensional fusion feature tensor. The weight distribution map is multiplied element-wise with the input high-dimensional fusion feature tensor to obtain the weighted high-dimensional fusion feature tensor.

8. The comprehensive performance testing system for precision forged motor shafts and gear sets according to claim 7, characterized in that: The specific process by which the multi-task prediction network head outputs three prediction results and their corresponding cognitive uncertainty estimates is as follows: The multi-task prediction network head contains a shared low-level feature extraction layer and branches into three independent sub-networks at higher levels, corresponding to the defect existence classification, defect type classification, and defect severity level regression tasks, respectively. The defect existence classification subnetwork outputs a scalar between zero and one, representing the probability that the detected workpiece has a defect corresponding to the weighted high-dimensional fused feature tensor; the defect type classification subnetwork outputs a multi-dimensional vector, the dimension of which is equal to the total number of predefined defect categories, and the value of each element in the multi-dimensional vector represents the probability that the detected workpiece belongs to the corresponding defect category; the defect severity regression subnetwork outputs a scalar greater than or equal to zero, representing the numerical value of the defect severity of the detected workpiece. During the inference phase, the random dropout layer in the network is kept active, and multiple independent forward propagation calculations are performed on the same weighted high-dimensional fusion feature tensor. Each calculation yields a set of prediction results consisting of the probability of defect existence, the probability distribution of defect type, and the numerical value of defect severity level due to the introduction of random dropout. For the defect existence classification task, the cognitive uncertainty is estimated by calculating the average information entropy of the defect existence probability values ​​obtained from multiple predictions. That is, firstly, based on the defect existence probability value obtained from each forward propagation, the information entropy corresponding to that prediction is calculated. This information entropy is equal to the negative defect existence probability value multiplied by the natural logarithm of the defect existence probability value, and the difference between the negative one minus the defect existence probability value multiplied by the natural logarithm of the difference, and the sum of the two. Then, the arithmetic mean of the information entropy calculated from all forward propagations is obtained. For the defect type classification task, the cognitive uncertainty is estimated by calculating the average information entropy of the type probability distribution obtained from multiple predictions. That is, firstly, based on the defect type probability distribution vector obtained from each forward propagation, the information entropy corresponding to the prediction is calculated. This information entropy is equal to the sum of the products of each probability value in the negative defect type probability distribution vector and its natural logarithm. Then, the arithmetic mean of the information entropy calculated from all forward propagations is obtained. For a severity level regression task, the cognitive uncertainty is estimated by calculating the standardized interquartile range (IIR) of the severity levels obtained from multiple predictions. The specific process is as follows: First, the severity levels obtained from multiple predictions are sorted in ascending order; then, the upper and lower quartiles of this sorted sequence are found; next, the difference between the upper and lower quartiles is calculated; then, the sum of the upper and lower quartiles is calculated; finally, the difference is divided by the sum and a positive constant; the quotient is the standardized IIR, which serves as the estimate of the cognitive uncertainty of the regression task. The multi-task prediction network head outputs in parallel the final probability of defect existence, the final probability distribution of defect type, the final regression value of defect severity level, and the corresponding three cognitive uncertainty estimates.

9. The comprehensive performance testing system for precision forged motor shafts and gear sets according to claim 8, characterized in that: In the model evolution module, the specific process of marking samples with cognitive uncertainty estimates higher than a preset threshold as fuzzy samples and storing them in a buffer queue is as follows: First, the cognitive uncertainty estimates for defect existence classification, defect type classification, and defect severity level regression output by the intelligent decision-making module are processed by nonlinear transformation. Specifically, the cognitive uncertainty estimates for defect existence classification and defect type classification are processed using a first nonlinear function, which is in the form of an exponential function that is a negative scaling factor with the natural constant e as the base multiplied by the input value. The cognitive uncertainty estimation of the defect severity level regression is processed using a second nonlinear function, which is in the form of: constant 2 divided by pi, then multiplied by the input value multiplied by the scaling factor, resulting in the arctangent value. Next, for the cognitive uncertainty estimates of defect existence classification, defect type classification, and defect severity regression after processing by the first nonlinear function, a first weight coefficient, a second weight coefficient, and a third weight coefficient are assigned, respectively. The first weight coefficient is multiplied by the cognitive uncertainty estimate of defect existence classification after processing by the first nonlinear function to obtain the first product; the second weight coefficient is multiplied by the cognitive uncertainty estimate of defect type classification after processing to obtain the second product; the third weight coefficient is multiplied by the cognitive uncertainty estimate of defect severity regression after processing to obtain the third product; the first, second, and third products are added together to obtain a comprehensive uncertainty score. Then, a dynamic threshold is calculated as follows: multiply the smoothing coefficient by the average comprehensive uncertainty score of the historical buffer queue to obtain the first term; subtract the smoothing coefficient from one to obtain a difference coefficient; calculate the larger value between the average comprehensive uncertainty score of the current buffer queue and a preset minimum threshold; multiply the difference coefficient by the larger value to obtain the second term; finally, add the first term and the second term together, and the sum is the dynamic threshold. Finally, the comprehensive uncertainty score corresponding to the data currently being processed and output by the intelligent decision module is compared with the dynamic threshold. If the comprehensive uncertainty score is greater than the dynamic threshold, this set of data and its corresponding detection object are marked as fuzzy samples, and they, along with the corresponding original acoustic vibration signal, optical size measurement value, background acoustic feature vector and prediction results from the intelligent decision module, are stored in the buffer queue.

10. The comprehensive performance testing system for precision-forged motor shafts and gear sets according to claim 9, characterized in that: The specific process of incrementally learning and updating the network model parameters in the data synthesis module, the feature extraction module, and the intelligent decision-making module using newly labeled samples is as follows: First, the newly labeled samples with precise labels obtained through the human-machine collaborative labeling process are processed to generate soft labels for the classification task. The soft label is a weighted combination of the precise label and the original prediction result of the intelligent decision module for the sample, where the weight of the precise label is higher than the weight of the original prediction result. The specific method for generating soft labels is as follows: For the defect type classification task, the defect type in the precise label is converted into a one-hot encoded vector, the probability distribution vector of the defect type output by the intelligent decision module for the sample is multiplied by a preset distillation coefficient, the one-hot encoded vector is multiplied by the difference between the one-hot and distillation coefficients, and the two product results are added together. The sum is the soft label used for training. Subsequently, based on the newly labeled samples and their soft labels, the network model parameters in the feature extraction module and the intelligent decision-making module are updated. The loss function used for the update consists of two parts: the first part is the standard task loss on the newly labeled samples, and the second part is a penalty term for each trainable parameter in the network model. The penalty term is calculated as follows: for each trainable parameter in the network model, a non-negative importance metric representing the importance of the parameter to the learned historical tasks is obtained, and the importance metric, a preset global regularization strength coefficient, and the square of the difference between the current value of the parameter and the old value saved before the start of this incremental learning are multiplied together. The resulting product is the penalty term for the parameter. The penalty terms for all trainable parameters are summed to obtain the second part of the loss function. Meanwhile, the conditional generative adversarial network in the data synthesis module is updated while keeping its discriminator parameters unchanged. The generator is fine-tuned using only real defect samples and their labels from the newly labeled samples. During the fine-tuning process, the optimization objective of the generator is to make the conditions of its generated synthetic acoustic and vibration signals match the defect types and severity level labels in the newly labeled samples, so that it can learn to synthesize such newly confirmed defect patterns.

Images