A Deep Learning-Based Intelligent Detection System for Gear Shaft Machining Defects

Abstract

This invention discloses a deep learning-based intelligent detection system for gear shaft machining defects, belonging to the field of intelligent manufacturing and quality control technology. It includes a multimodal sensor data acquisition module for acquiring multimodal time-series signals during machining to generate a raw sensor data stream; a process-morphology cross-scale mapping module for receiving the raw sensor data stream and mapping it into a deep causal state feature sequence; and a defect generation prediction and tracing module for combining the deep causal state feature sequence with the tool's instantaneous spatial coordinates to calculate the conditional probability of specific types of defects and construct a defect probability spatial distribution map covering the workpiece's machined surface. This invention constructs a complete closed-loop intelligent control system, solving the fundamental problem of existing technologies that can only analyze real-time process data but cannot accurately predict the specific morphology and location of defects before they form, thus greatly improving the yield of high-precision gear shaft machining.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and quality control technology, specifically to an intelligent detection system for gear shaft machining defects based on deep learning. Background Technology

[0002] In the automated machining of high-precision gear shafts, controlling the surface quality of the product is a core aspect. Traditional quality control methods are typically divided into two independent categories: process monitoring and post-processing inspection. Process monitoring systems determine whether the machining state is abnormal by analyzing the statistical characteristics of sensor signals such as cutting force and vibration, but they cannot predict what specific surface defects the abnormality will lead to. Post-processing inspection, on the other hand, uses optical or non-destructive testing equipment to identify and classify defects after machining is completed. By this time, defective products have already been produced, resulting in a waste of materials and time. Existing technologies have failed to establish a deterministic causal relationship between dynamic sensor data during the machining process and the final microscopic defect morphology, thus making it impossible to provide effective early warning and intervention before defects form. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent detection system for gear shaft machining defects based on deep learning, so as to solve the problems mentioned in the background art.

[0004] The technical solution of this invention includes a multimodal sensing data acquisition module for acquiring multimodal time-series signals during the processing to generate a raw sensing data stream;

[0005] The process-topography cross-scale mapping module is used to receive the raw sensor data stream and map it into a deep causal state feature sequence;

[0006] The defect generation prediction and tracing module is used to combine the deep causal state feature sequence with the tool's instantaneous spatial coordinates to calculate the conditional probability of a specific type of defect and construct a spatial distribution map of the defect probability covering the workpiece's machined surface.

[0007] The adaptive control module for processing technology is used to compare the defect probability in the defect probability spatial distribution map with a preset risk threshold. When the defect probability exceeds the risk threshold, a control command is generated and executed. When the defect probability does not exceed the risk threshold, the current processing technology is maintained.

[0008] Preferably, the multimodal time series signal includes cutting force signal, triaxial vibration acceleration signal and acoustic emission signal.

[0009] Preferably, the specific processing steps of the process-morphology cross-scale mapping module are as follows:

[0010] Calculate the instantaneous spectral entropy of each type of signal in the raw sensor data stream;

[0011] Calculate the normalized cross-correlation function values ​​between different physical signals;

[0012] By combining the preset entropy weight coefficient and cross-correlation weight coefficient, the instantaneous spectral entropy and the normalized cross-correlation function value are weighted and summed to generate a deep causal state feature sequence.

[0013] Preferably, the entropy weight coefficient and the cross-correlation weight coefficient are determined by principal component analysis or sensitivity analysis of historical data.

[0014] The preferred steps for calculating conditional probability are as follows:

[0015] Based on the deep causal state feature sequence under historical processing states, exponential decay weighting is performed and integrated along the time dimension to calculate the weighted cumulative damage value of historical states.

[0016] The weighted cumulative damage value is input into the Sigmoid activation function for mapping to generate conditional probabilities.

[0017] Preferably, the sensitivity rate weights, forgetting factors, and bias parameters used in the calculation steps are obtained by training a deep learning model on a labeled dataset containing process data and 3D surface topography scan data.

[0018] Preferably, before generating control instructions, the adaptive control module for processing technology is also used for:

[0019] Analyze the moment when the deep causal state feature contributes the most to the increase in defect probability.

[0020] The original sensor data corresponding to the extracted time is analyzed for spectral characteristics to generate defect source information that identifies the cause of the defect.

[0021] Preferably, the control command is generated based on the specific frequency vibration identified in the defect tracing information, and is used to fine-tune the spindle speed to avoid the resonance zone.

[0022] Preferably, the risk threshold is determined by analyzing the receiver operating characteristic curves of historical data and balancing the cost of defective products with the loss of production efficiency.

[0023] This invention provides an intelligent detection system for gear shaft machining defects based on deep learning, the advantages of which are:

[0024] This invention constructs a complete closed-loop intelligent control system, including a multimodal sensor data acquisition module, a process-morphology cross-scale mapping module, a defect generation prediction and tracing module, and a machining process adaptive control module. This system can monitor the machining process in real time, accurately predict the type and spatial location of defects on the workpiece surface, and actively control the process based on the cause of defects, thereby eliminating defects before they form. This solves the fundamental problem of existing technologies that can only analyze real-time process data but cannot accurately predict the specific morphology and location of defects before they form, greatly improving the machining yield of high-precision gear shafts.

[0025] The multimodal sensing data acquisition module of this invention obtains physical information that can comprehensively characterize the dynamic interaction state of the tool-workpiece-machine tool system by simultaneously acquiring cutting force signals, triaxial vibration acceleration signals and acoustic emission signals. Compared with single signal monitoring, this combination of multimodal data with complementary physical connotations provides richer and more robust original evidence for subsequent feature extraction, significantly enhancing the system's ability to perceive complex machining states and the reliability of defect prediction.

[0026] The process-morphology cross-scale mapping module of this invention can extract a low-dimensional, highly information-dense deep causal state feature sequence from high-dimensional, noisy raw sensor data streams. This module objectively and accurately reveals the inherent risk of the processing system tending to form surface micro-defects by structurally quantifying the instantaneous spectral entropy of each signal and the normalized cross-correlation function between different signals, and by combining the entropy weight coefficients and cross-correlation weight coefficients determined by principal component analysis or sensitivity analysis for weighted summation. This data-driven feature generation method avoids the subjectivity of manually setting parameters, enabling the feature to more effectively characterize the deterioration trend of the processing state.

[0027] The defect generation prediction and tracing module of this invention establishes a probabilistic model that simulates the physical process of damage accumulation, and associates the deep causal state feature sequence in the time dimension with the instantaneous spatial coordinates of the tool. This module performs exponential decay weighting on historical states and integrates along the time dimension to calculate the weighted cumulative damage value of historical states, and maps it to the conditional probability of a specific type of defect. This method not only accurately simulates the physical nature of defect formation, but also constructs a defect probability space distribution map covering the entire workpiece machining surface. The key sensitivity rate weights, forgetting factors, and bias term parameters in the model are all obtained through deep learning model training, ensuring high accuracy and strong adaptability of the prediction results.

[0028] The adaptive control module for processing technology of this invention achieves precise and efficient closed-loop intervention. Before generating control commands, the module can analyze the moment corresponding to the deep causal state characteristics that contribute the most to the defect probability increase stage, and perform spectral characteristic analysis on the original sensor data at that moment to generate defect source information that identifies the cause of the defect. Based on this, the system can generate targeted control commands, such as fine-tuning the spindle speed according to the identified specific frequency vibration to avoid the resonance zone. This source-based precise control achieves targeted solutions, efficiently solving problems with minimal process adjustment, and minimizing the negative impact on production efficiency.

[0029] The risk threshold used for decision-making in this invention is scientifically determined by analyzing the receiver operation characteristic curves of historical data and balancing the cost of defective products and the loss of production efficiency. This systematic setting method overcomes the subjectivity and non-optimality of traditional experience-based threshold setting, enabling the system's decision-making behavior to reach the optimal balance point according to the economic requirements of production, thereby significantly improving the overall economic efficiency of production. Attached Figure Description

[0030] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0031] Figure 1 This is a flowchart of an intelligent detection system for gear shaft machining defects based on deep learning, according to the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0033] Example 1:

[0034] Please see Figure 1 This invention provides a deep learning-based intelligent detection system for gear shaft machining defects, comprising:

[0035] The multimodal sensing data acquisition module is used to acquire multimodal time-series signals during the processing to generate raw sensing data streams;

[0036] The process-topography cross-scale mapping module is used to receive the raw sensor data stream and map it into a deep causal state feature sequence;

[0037] The defect generation prediction and tracing module is used to combine the deep causal state feature sequence with the tool's instantaneous spatial coordinates to calculate the conditional probability of a specific type of defect and construct a spatial distribution map of the defect probability covering the workpiece's machined surface.

[0038] The adaptive control module for processing technology is used to compare the defect probability in the defect probability spatial distribution map with a preset risk threshold. When the defect probability exceeds the risk threshold, a control command is generated and executed. When the defect probability does not exceed the risk threshold, the current processing technology is maintained.

[0039] A multimodal sensing data acquisition module is deployed on the gear shaft machining equipment. The purpose of this module is to capture physical signals that can comprehensively reflect the dynamic interaction state of the tool-workpiece-machine tool system in real time and synchronously. In this embodiment, the module acquires multimodal time series signals at high frequency by installing various types of sensors at key positions of the machine tool. These signals are integrated to form a raw sensing data stream and transmitted to the subsequent processing module in real time.

[0040] The process-morphology cross-scale mapping module receives the raw sensor data stream; the purpose of this module is to extract a low-dimensional but higher information-density feature from the high-dimensional, noisy raw sensor data. This feature can directly characterize the inherent risk of the processing system tending to form surface micro-defects. In this embodiment, this module calculates a deep causal state feature. To achieve this goal; deep causal state characteristics This refers to a single time-series scalar that quantifies the complexity and coupling of a system under multiphysics conditions. Its function is to reduce the dimensionality of data and make nonlinear modes such as resonance or instability, which are directly related to defects, explicit. The formula for calculating this feature is defined as follows:

[0041]

[0042] The parameters in the formula are defined as follows:

[0043] Summation is performed for all sensor signal types;

[0044] To sum all different types of sensor signal pairs;

[0045] The output of this module at time The deep causal state characteristic is a dimensionless scalar, and its numerical fluctuation characterizes the degree of risk that the processing system tends to form defects.

[0046] This serves as an index for the type of sensor signal, such as representing cutting force, X-axis vibration, etc.

[0047] For the first The entropy weight coefficient of the signal type is a dimensionless scalar. This parameter is determined by principal component analysis or sensitivity analysis of historical data and reflects the contribution of the complexity of different types of signals to the formation of defects.

[0048] For the first in the original sensing data stream The signal calculated at time 10:00 The instantaneous spectral entropy is a dimensionless value. This value is obtained by calculating the information entropy of the normalized power spectral density within a certain time window after performing a short-time Fourier transform on the signal, and it reflects the frequency complexity of the signal at the current moment.

[0049] For signal and The cross-correlation weighting coefficient between them is a dimensionless scalar; the source of this parameter is related to... Similarly, historical data analysis is used to determine the contribution of the coupling effect between specific signal pairs to defect formation;

[0050] For the signal in the original sensing data stream and The calculated time delay The normalized cross-correlation function value is a dimensionless value used to quantify the coupling strength and response delay between two different physical signals;

[0051] The time delay parameter is determined by analyzing the sequence of different signal responses in historical data to capture the causal lag effect in the physical process.

[0052] Time delay parameter It is determined by calculating the peak value of the cross-correlation function between different signal pairs at different delays, and its value is taken as the delay that makes the cross-correlation function reach its maximum value;

[0053] Calculated The time series data is used as direct input to the subsequent prediction module;

[0054] The defect generation prediction and tracing module receives the deep causal state feature sequence generated by the previous module. It is the time from the start of processing. up to the current moment The module aims to establish a deterministic mapping relationship between the state evolution in the time dimension of the processing process and the defects ultimately formed in the spatial dimension of the workpiece, thereby achieving online defect prediction. In this embodiment, the module obtains the state of the CNC system of the processing equipment in real time, thus achieving online defect prediction. Instantaneous spatial coordinates of the tool acting on the workpiece surface It utilizes an integral probability model with a forgetting factor to incorporate time features. With instantaneous spatial coordinates In essence, it simulates the cumulative damage effect in physics; this model is used to calculate the damage at the current processing location. A specific type appears at the location The conditional probability of a defect is expressed mathematically as follows:

[0055]

[0056] The parameters in the formula are defined as follows:

[0057] Let be the conditional probability, a dimensionless value representing the probability of success at the current tool position given a historical machining state. The first one appeared at the place The possibility of class defects (where, Represents appearance, Representative category, (Represents the position vector)

[0058] The sigmoid activation function maps the internal cumulative damage value to... The probability range;

[0059] In order to target the The sensitivity rate weight for class defects, whose physical dimension is the reciprocal of time. This parameter is obtained through training a deep learning model, and it quantifies the causal state characteristics per unit time. The strength of the contribution to the logarithmic occurrence rate of a specific defect type;

[0060] The forgetting factor, also known as the decay rate, has the physical dimension of the reciprocal of time. This parameter, also obtained from model training, determines the rate at which historical state information decays, reflecting the duration of the processing system's memory of past disturbances.

[0061] It is an exponentially decaying kernel function, dimensionless, which gives higher weight to recent states than to distant states;

[0062] The value calculated by the preprocess-morphological cross-scale mapping module at a certain point in the past. The deep causal state characteristics;

[0063] Let be the bias term, and be a dimensionless learnable parameter representing the bias term in the absence of any process perturbation. The inherent tendency of class defects to occur;

[0064] As the processing progresses, this module will process each calculated probability value. Its corresponding spatial coordinates By performing correlation, a complete, high-resolution spatial distribution map of defect probability covering the workpiece's machined surface is finally constructed;

[0065] The adaptive processing control module receives the defect probability spatial distribution map output by the prediction module. The purpose of this module is to proactively and intelligently adjust processing parameters before defects actually form, based on the predicted defect risks, to avoid or mitigate defect occurrence in a closed-loop manner. In this embodiment, the module will generate the defect probability in real time. The probability of a defect is compared with a preset risk threshold. The risk threshold is a decision boundary used to determine whether the probability of a defect occurs reaches a level that requires intervention. When the probability of a defect exceeds the risk threshold, the module will automatically generate and execute control instructions. When the probability of a defect does not exceed the risk threshold, the current processing technology will be maintained to avoid unnecessary intervention that could affect production efficiency.

[0066] The technical effect of the system disclosed in this embodiment is that by constructing a complete mapping and prediction link from multimodal sensing data to deep causal state features and then to the spatial probability distribution of defects, and combining it with closed-loop adaptive control, it solves the fundamental problem that existing technologies cannot accurately predict the type, size and location of defects by analyzing real-time process data before defects are formed. It transforms traditional post-processing inspection into pre-processing prediction and intervention, which can eliminate potential defects before defective products are produced, thereby greatly improving the yield of high-precision gear shaft machining, reducing the waste of materials and time, and realizing truly intelligent quality control.

[0067] Example 2:

[0068] Multimodal time series signals include cutting force signals, triaxial vibration acceleration signals, and acoustic emission signals;

[0069] Based on the system in Example 1, the specific content acquired by the multimodal sensing data acquisition module is defined; in this example, the multimodal time series signal is specified as including at least three key physical signals: cutting force signal, triaxial vibration acceleration signal, and acoustic emission signal;

[0070] The cutting force signal is acquired by a force sensor installed on the tool holder or spindle. It reflects the interaction force between the tool and the workpiece during the cutting process and its function is to directly characterize the cutting load and tool state.

[0071] The triaxial vibration acceleration signal is a vibration signal in the X, Y, and Z directions collected by accelerometers installed on key components of the machine tool (such as the spindle box and worktable). Its function is to comprehensively monitor the dynamic stability of the machine tool system.

[0072] Acoustic emission signals are transient elastic wave signals released inside the material due to the generation or propagation of tiny cracks, which are collected by acoustic sensors installed on the workpiece or fixture near the cutting area. They are highly sensitive to microscopic damage events inside the material.

[0073] The gain technology achieves its effect by simultaneously acquiring three signals from different sources with complementary physical meanings, enabling a more comprehensive and robust capture of the state information of the machining system. Cutting force directly reflects the machining load, vibration signal reflects system stability, while acoustic emission signal can reveal early damage at the material level. This multimodal combination, compared to single signal monitoring, can provide richer and more discriminative raw information for subsequent deep causal state feature extraction, thereby improving the accuracy and reliability of defect prediction.

[0074] Example 3:

[0075] The specific processing steps of the process-topography cross-scale mapping module are as follows:

[0076] Calculate the instantaneous spectral entropy of each type of signal in the raw sensor data stream;

[0077] Calculate the normalized cross-correlation function values ​​between different physical signals;

[0078] By combining preset entropy weight coefficients and cross-correlation weight coefficients, the instantaneous spectral entropy and normalized cross-correlation function values ​​are weighted and summed to generate a deep causal state feature sequence. This invention integrates the dynamic interaction information of multiple physics fields within the system by weighted fusion of the instantaneous frequency complexity and coupling strength between signals, thus characterizing the evolution trend of the processing state from a higher dimension. The depth of this feature lies in its data-driven weight determination method, which reveals the nonlinear mapping relationship between the original signal and the final defect. Its causality is reflected in its ability to capture the dynamic hysteresis effect during the processing.

[0079] The entropy weight coefficient and cross-correlation weight coefficient are determined by principal component analysis or sensitivity analysis of historical data;

[0080] Based on the system in Example 1, the specific calculation steps for generating deep causal state feature sequences by the process-morphology cross-scale mapping module are described in detail. The specific processing procedure of this module is as follows: First, each type of signal in the original sensing data stream is processed to calculate its instantaneous spectral entropy. The underlying logic lies in quantifying the frequency complexity or uncertainty of each type of signal at the current moment. The more chaotic and disordered the signal spectrum, the higher its entropy value, which is usually related to the instability of the processing state. Secondly, it calculates the normalized cross-correlation function values ​​between different physical signal pairs. Its purpose is to quantify the coupling strength and response timing relationship between different physical phenomena; based on the above calculation results, the module will calculate the instantaneous spectral entropy. and cross-correlation function values Multiply by their respective weighting coefficients and They are then summed and ultimately merged into a comprehensive state index, namely, a deep causal state characteristic. ;

[0081] To ensure the objectivity and effectiveness of this fusion process, the entropy weight coefficient is... cross-correlation weight coefficient It is determined by principal component analysis or sensitivity analysis on historical data; principal component analysis or sensitivity analysis refers to a data-driven parameter optimization method. Its working principle is to analyze the spectral entropy of each signal and the cross-correlation between signals and the correlation strength of the final defect formation on a historical dataset containing a large amount of processing data and surface quality annotations of the final product, so as to learn the optimal combination of weight coefficients.

[0082] The gain technique, through its step-by-step computation and weighted fusion approach, not only structurally extracts deep information from both signal complexity and inter-signal coupling dimensions, but also determines weight coefficients in a data-driven manner, avoiding the subjectivity and blindness of manually setting parameters. This results in the final generated deep causal state features. It can reflect the deterioration trend of the processing status more objectively and accurately, laying a solid data foundation for subsequent accurate prediction, and improving the intelligence level of the entire system and the generalization ability of the model.

[0083] Example 4:

[0084] The steps for calculating conditional probability are as follows:

[0085] Based on the deep causal state feature sequence under historical processing states, exponential decay weighting is performed and integrated along the time dimension to calculate the weighted cumulative damage value of historical states.

[0086] The weighted cumulative damage value is input into the Sigmoid activation function for mapping to generate conditional probabilities;

[0087] The sensitivity rate weights, forgetting factors, and bias parameters used in the calculation steps were obtained by training a deep learning model on a labeled dataset containing process data and 3D surface topography scan data.

[0088] Based on the system in Example 1, the specific steps for calculating the conditional probability of the defect generation prediction and tracing module are explained in detail. This calculation process first performs exponential decay weighting on the deep causal state feature sequence under historical processing states and integrates it along the time dimension. The underlying mechanism is that for each historical moment in the sequence… eigenvalues Using the exponentially decaying kernel function We weight the features so that more recent states have a greater impact on the current state; then, we multiply the weighted feature values ​​by the sensitivity rate weight. Integrate along the time axis; the result of this integration is compared with a bias term. The summations together constitute the weighted cumulative damage value of the historical state; this step aims to simulate the physical process of defect formation, i.e., the final macroscopic defect is the result of the continuous accumulation of historical microscopic damage; then, the weighted cumulative damage value calculated in the previous step is input into the Sigmoid activation function for mapping, which transforms an internal, physically significant damage level into a standard, localized damage value. Conditional probability within an interval ;

[0089] To further clarify, the three core parameters used in the calculation steps are the sensitivity rate weight. Forgetting factors With bias term It is obtained by training a deep learning model on a labeled dataset containing process data and 3D surface topography scanning data; deep learning model training refers to a process in which a large amount of historical processing sensor data is used as input, and the corresponding true types and locations of workpiece surface defects obtained by a high-precision 3D scanner are used as labels. Through optimization algorithms such as backpropagation, the model parameters are automatically adjusted. This minimizes the difference between the defect probability distribution predicted by the model and the actual defect distribution.

[0090] This integral probability model is implemented using a recurrent neural network or a long short-term memory network; the network uses a deep causal state feature sequence. As input, its internal state is responsible for simulating the cumulative damage value, and finally outputs the conditional probability through the Sigmoid activation function; in the formula... These are the network's weights, decay rates, and bias parameters, obtained through end-to-end training on the labeled dataset.

[0091] The deep learning model in this invention is a multilayer perceptron or recurrent neural network, trained using the Adam optimizer and binary cross-entropy loss function; the training dataset is a labeled dataset containing historical processing sensing data and corresponding three-dimensional surface topography scanning data; the training objective is to minimize the difference between the defect probability distribution predicted by the model and the actual defect distribution.

[0092] The gain technique enables this integral-based damage accumulation model to accurately simulate the physical nature of defect formation, making the prediction results more interpretable. More importantly, by obtaining model parameters through end-to-end deep learning training, the model can automatically learn from the data the sensitivity of different types of defects to process disturbances, the system's memory effect, and the inherent occurrence rate, which greatly improves the accuracy of the prediction model and its adaptability to different processing conditions, achieving highly customized and automated defect prediction.

[0093] Example 5:

[0094] Before generating control commands, the adaptive control module for machining processes is also used for:

[0095] Analyze the moment when the deep causal state feature contributes the most to the increase in defect probability.

[0096] The original sensor data corresponding to the time point is extracted and its spectral characteristics are analyzed to generate defect source information that identifies the cause of the defect.

[0097] The control command is generated based on the specific frequency vibration identified in the defect tracing information, and is used to fine-tune the spindle speed to avoid the resonance zone.

[0098] The risk threshold is determined by analyzing the receiver's operational characteristic curves of historical data, after balancing the cost of defective products with the loss of production efficiency.

[0099] Based on the system in Example 1, the defect tracing function of the adaptive control module for processing technology before generating control commands is described in detail; when the module detects the probability of a defect at a certain location... When a risk begins to rise significantly and shows a tendency to exceed the risk threshold, a source analysis will be performed before generating regulatory instructions. This analysis process includes back-calculating the integral of the probability value and identifying the historical moments that made the major contribution to the integral term. These moments are considered critical points that lead to an increased risk of current defects; subsequently, the module uses the identified critical moments... The system precisely retrieves signal segments before and after a given moment from the stored raw data stream and performs detailed spectral analysis on them to identify any abnormal frequency components. The analysis results are then integrated into defect tracing information that identifies the cause of the defect.

[0100] In a specific application scenario of this embodiment, the control command is generated based on the specific frequency vibration identified in the defect tracing information. It is used to fine-tune the spindle speed to avoid the resonance zone. For example, if the tracing information shows that the main reason for the increase in the defect probability is the surge in vibration energy at a specific frequency, and this frequency is close to a certain harmonic of the current spindle speed, the system can determine that cutting resonance has occurred. Accordingly, the control module will automatically generate a command and send it to the CNC system of the machine tool. The command content is to fine-tune the spindle speed, thereby changing the excitation frequency of the system and actively destroying the resonance condition.

[0101] Based on the system in Example 1, the risk threshold in the adaptive control module for processing technology is adjusted. The setting method is described in detail; the risk threshold is described in detail. It is not a fixed value set based on experience, but rather the optimal decision point determined by analyzing the receiver's operational characteristic curve of historical data and comprehensively considering the defective product costs caused by underreporting and the production efficiency losses caused by false reporting.

[0102] ROC curve analysis is a statistical method for evaluating the performance of binary classifiers. In this embodiment, its application logic is as follows: First, on a dataset containing a large number of historical processed samples, by iterating through all possible probability values ​​as thresholds, the true positive rate (the proportion of correctly predicted defects) and false positive rate (the proportion of incorrectly classified defects) are calculated at each threshold. Then, an ROC curve is plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. Finally, a cost-benefit analysis is introduced to quantify the costs of missed and false positives, and combined with the true positive and false positive rates at each point on the ROC curve, the total expected cost at each threshold is calculated. The point that minimizes the total expected cost is selected, and its corresponding probability value is determined as the optimal risk threshold. ;

[0103] The gain technology achieves its effect by determining risk thresholds in a systematic and data-driven manner, overcoming the subjectivity and non-optimal nature of traditional methods that rely on human experience to set thresholds. This method can scientifically find an optimal balance point based on the company's specific trade-offs between quality costs and production efficiency. This optimizes the system's decision-making behavior, ensuring that while effectively preventing the generation of defective products, unnecessary interference with normal production is minimized, thereby improving the overall economic efficiency of production.

[0104] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based gear shaft machining defect intelligent detection system, characterized in that, The method comprises the following steps: a multi-modal sensor data acquisition module is used to collect multi-modal time series signals in the machining process to generate raw sensor data streams; a process-topography cross-scale mapping module is used to receive the raw sensor data streams and map them into deep causal state feature sequences; a defect generative prediction and tracing module is used to combine the deep causal state feature sequences with the instantaneous spatial coordinates of the tool to calculate the conditional probability of a specific type of defect and construct a defect probability spatial distribution map covering the machined surface of the workpiece; a machining process self-adaptive regulation and control module is used to compare the defect probability in the defect probability spatial distribution map with a preset risk threshold value, and when the defect probability exceeds the risk threshold value, a regulation and control instruction is generated and executed, and when the defect probability does not exceed the risk threshold value, the current machining process is maintained; the specific processing steps of the process-topography cross-scale mapping module are as follows: the instantaneous spectral entropy of each type of signal in the raw sensor data stream is calculated; the normalized cross-correlation function value between different physical signals is calculated; the instantaneous spectral entropy and the normalized cross-correlation function value are weighted and summed by combining the preset entropy weight coefficient and the cross-correlation weight coefficient to generate the deep causal state feature sequence; the multi-modal time series signals include cutting force signals, three-axis vibration acceleration signals and acoustic emission signals; the entropy weight coefficient and the cross-correlation weight coefficient are determined by principal component analysis or sensitivity analysis on historical data; the calculation steps of the conditional probability are as follows: based on the deep causal state feature sequences under the historical machining state, exponential decay weighting is performed and integration is performed along the time dimension to calculate the weighted cumulative damage value of the historical state; the weighted cumulative damage value is input into a Sigmoid activation function for mapping to generate the conditional probability; the sensitivity rate weight, the forgetting factor and the bias term parameters used in the calculation steps are obtained by deep learning model training on a labeled data set containing process data and three-dimensional surface topography scanning data.

2. The deep learning-based gear shaft machining defect intelligent detection system according to claim 1, wherein Before generating the regulation and control instruction, the machining process self-adaptive regulation and control module is also used to: analyze the time corresponding to the deep causal state feature that contributes most to the rising stage of the defect probability; extract the raw sensor data corresponding to the time for spectral characteristic analysis to generate defect tracing information that identifies the causes of the defect.

3. The deep learning-based gear shaft machining defect intelligent detection system according to claim 2, characterized in that, The regulation and control instruction is an instruction for fine-tuning the spindle speed to avoid the resonance region according to the specific frequency vibration identified in the defect tracing information.

4. The deep learning-based gear shaft machining defect intelligent detection system according to claim 1, wherein The risk threshold value is determined by analyzing the receiver operating characteristic curve of the historical data after balancing the cost of defective products and the loss of production efficiency.

Images