Ocean engineering gear machining real-time quality prediction system based on multi-sensor fusion

Abstract

The invention discloses a multi-sensor fusion ocean engineering gear machining real-time quality prediction system, which relates to the technical field of ocean engineering gear machining prediction and comprises a multi-dimensional sensing data acquisition module, a self-adaptive data preprocessing module, a spatial-temporal feature fusion module, a quality prediction model module, a real-time decision feedback module and a system cooperative control module. The acquisition module comprises multiple types of sensors and synchronously records process parameters; the preprocessing module is used for reducing noise, smoothing data and complementing abnormal values; a feature fusion module extracts multi-source data feature construction vectors; the prediction model outputs a quality index by using an improved DBN; the decision-making module is used for comparing threshold early warning or adjusting parameters; and the collaboration module adopts an edge-cloud architecture to guarantee real-time prediction and model updating. The multi-sensor fusion covers the full-dimension information, and the quality is predicted in real time, so that the waste of post detection is avoided; the model is high in adaptability and can cope with multi-material working conditions; defect sources are positioned accurately, and machining quality and production efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of marine engineering gear processing prediction technology, and in particular to a real-time quality prediction system for marine engineering gear processing based on multi-sensor fusion. Background Technology

[0002] As core transmission components of marine platforms and ship propulsion systems, marine engineering gears must withstand harsh conditions such as high loads, high salt spray, and alternating impacts over extended periods. Their machining quality directly determines the operational safety and service life of marine engineering equipment. Currently, marine engineering gear machining often employs a "post-production inspection" model, where gears are inspected after machining using equipment such as coordinate measuring machines and surface roughness testers. This model suffers from significant delays. If defects such as excessive tooth profile accuracy or insufficient surface hardness are found during inspection, the machined gears must be reworked or even scrapped. This not only wastes raw materials and machining time but may also delay the assembly cycle of marine engineering equipment and increase production costs.

[0003] With the development of intelligent manufacturing technology, some enterprises have attempted to introduce single sensors to monitor the gear machining process. However, due to the limited monitoring dimensions, it is difficult to comprehensively reflect the machining quality status. For example, monitoring cutting vibration solely through vibration sensors can determine tool wear, but it cannot correlate changes in tooth surface roughness and surface hardness. Relying on temperature sensors can only control the impact of cutting heat on gear deformation, but cannot capture subtle changes in tooth profile dimensional deviations. Furthermore, marine engineering gear machining involves the coordinated control of multiple process parameters. Data from a single sensor cannot establish a mapping relationship between process parameters and quality indicators, making it difficult to quickly pinpoint the root cause when quality anomalies occur. Process adjustments rely on experience, resulting in low efficiency.

[0004] Furthermore, existing quality prediction models are mostly trained on samples of specific materials and operating conditions, resulting in poor generalization ability. When the processing material is changed to high-strength alloy steel or the operating conditions change, the model's prediction accuracy drops significantly, requiring the collection of a large number of new samples for retraining, a process that can take several weeks. This cannot meet the diverse, small-batch production needs of marine engineering gears. Simultaneously, sensor data during processing suffers from noise interference and spatiotemporal asynchrony; directly inputting raw data into the model easily leads to increased prediction errors. Data transmission between edge devices and the cloud has not been specifically optimized, resulting in a large amount of redundant data consuming bandwidth and affecting real-time prediction response speed, making it difficult to meet the triple requirements of "real-time performance, accuracy, and adaptability" for quality prediction in marine engineering gear processing. Summary of the Invention

[0005] The present invention proposes a multi-sensor fusion real-time quality prediction system for marine engineering gear processing to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-sensor fusion real-time quality prediction system for marine engineering gear processing, comprising the following modules: The multi-dimensional sensing data acquisition module adopts a distributed sensor array, including acoustic emission sensors, triaxial acceleration sensors, infrared thermal imagers, and laser displacement sensors. The sensors are connected to the data aggregation unit via industrial Ethernet to collect data during gear processing. The data is packaged and stored as structured data frames according to timestamps. The adaptive data preprocessing module removes Gaussian white noise from the acquired data using a wavelet threshold denoising algorithm; smooths the tooth profile size data of the laser displacement sensor using Kalman filtering; performs non-uniformity correction and temperature calibration on the infrared thermal image; and completes missing data using an improved Lagrange interpolation method, replacing outliers with the median value in a sliding window. The spatiotemporal feature fusion module calculates kurtosis, peak factor, and energy entropy for acoustic emission signals; performs Fourier transform on vibration signals to extract characteristic frequency amplitudes; extracts the highest tooth surface temperature, temperature gradient, and temperature uniformity coefficient from temperature field data; calculates tooth thickness deviation, cumulative tooth pitch error, and tooth profile error from tooth profile size data; and dynamically allocates feature weights through an attention mechanism to construct a feature matrix. The quality prediction model module receives 28-36 dimensional fused feature vectors as input, and the hidden layers contain 2-4 layers of restricted Boltzmann machines with 48-80, 24-40, and 12-20 neurons respectively. The output layer outputs the predicted values ​​of tooth surface roughness, tooth profile accuracy, and surface hardness respectively. The model is trained using the Adam optimizer with the root mean square error as the loss function, and the validation set accuracy is calculated every 30-70 epochs during training. The real-time decision feedback module compares the predicted results with the preset quality thresholds: when a single indicator exceeds the threshold by less than 10%, an early warning signal is generated and the deviation value is displayed; when it exceeds the threshold by more than 10%, a process parameter adjustment command is triggered, and the feed rate and spindle speed are dynamically corrected through the PID controller. The system's collaborative control module is responsible for data preprocessing and prediction at the edge, while the cloud platform is responsible for offline model training and updates. The edge and cloud communicate via a 5G private network, with the edge uploading feature data and prediction results every 0.5-2 hours and the cloud pushing model update parameters every 12-48 hours.

[0007] Furthermore, it also includes a multi-sensor spatiotemporal synchronization calibration unit, which synchronizes sensor time via hardware trigger signals; spatial calibration employs a three-dimensional coordinate transformation algorithm, through... Transform the measurement points in each sensor coordinate system to the machine tool coordinate system, where Three-dimensional coordinates in the machine tool coordinate system. R represents the three-dimensional coordinates in the sensor coordinate system, R is a 3×3 rotation matrix, and T is a 3×1 translation vector; the synchronous calibration results are displayed in real time on the system monitoring interface, and a calibration prompt is issued when the deviation exceeds the limit.

[0008] Furthermore, it also includes a processing condition adaptive adjustment submodule, which dynamically optimizes process parameters based on the matching degree between the prediction results and the current operating conditions: calculating the operating condition matching degree. Where n is the number of process parameters, For parameter weights, The current parameter value. The reference parameter value is k, which is the attenuation coefficient. When M < 0.6-0.8, the parameter optimization algorithm is started to search for the optimal parameter combination within the preset parameter range.

[0009] Furthermore, in the adaptive data preprocessing module, considering the non-stationary characteristics of the acoustic emission signal, a variational mode decomposition algorithm is used to decompose the signal: through... Solve for the optimal modal components, where For a set of K modal components, For the set of center frequencies of each mode, is the Dirac function, * is the convolution operation, t is the time variable, and j is the imaginary unit; after decomposition, the modal components with the highest energy percentage (2-4) are selected to reconstruct the signal.

[0010] Furthermore, the spatiotemporal feature fusion module introduces a feature mutual information weighting mechanism to calculate the mutual information between any two features. ,in , These are two different features. Let be the joint probability density function. , The marginal probability density function is used; the mutual information value is normalized as the feature association strength. Features with an association strength < 0.2 are assigned a weight of 0.4-0.6 times, and features with an association strength ≥ 0.2 are assigned a weight of 0.9-1.1 times, thereby enhancing the targeting and effectiveness of feature fusion.

[0011] Furthermore, in the quality prediction model module, a transfer learning strategy is employed to optimize the model's adaptability to new material gear processing scenarios: the model trained based on low-carbon steel gear data is used as the source model, through... Construct the migration loss function, where For the target domain loss, The regularization coefficient is . For the source model weights, Set the weights for the target model; freeze the parameters of the first 1-3 layers of the source model's RBM, and train only the last RBM layer and the output layer.

[0012] Furthermore, in the real-time decision feedback module, a quality traceability cause-effect graph model is constructed, and a Bayesian network is used to calculate the contribution of each process parameter to quality defects. ,in For the i-th process parameter, For the j-th type of quality defect, This represents the conditional probability of a defect occurring when parameters are abnormal. This represents the prior probability of parameter anomalies. The marginal probability of a defect occurring; parameters with a contribution greater than 0.3-0.5 are marked as key influencing factors and adjustment schemes are given priority in the decision recommendations.

[0013] Furthermore, the system collaborative control module employs a dynamic data compression and transmission strategy to compress the feature data uploaded from the edge: through... To achieve compressed sensing, in which For compressed data, The random measurement matrix is ​​6×32-10×36, and X is the original eigenvector; the compression ratio is 3:1-5:1, and the reconstruction error is ≤6%.

[0014] Furthermore, it also includes an online monitoring unit for gear material properties, integrating an eddy current sensor and an ultrasonic flaw detector to collect real-time data on the hardness distribution and internal defects of the gear blank; through The input to the quality prediction model is modified, where H is the modified material hardness. α represents the initial hardness, E represents the ultrasonic echo energy, and α and β are correction coefficients.

[0015] Furthermore, an online model evaluation and update triggering mechanism is added to the quality prediction model module: the average absolute percentage error between the predicted value and the actual detected value is calculated every 8-12 gears processed. Where n = 8 - 12, This is the actual measured value. The values ​​are predicted. When MAPE > 4-6%, incremental training in the cloud is automatically triggered, using 40-60 newly collected samples to update the model parameters, while maintaining normal system operation during the update process.

[0016] Compared with existing technologies, the beneficial effects of this invention are: In terms of real-time performance and comprehensiveness of quality prediction, the system constructs a multi-sensor fusion acquisition architecture. Through collaborative monitoring of multiple types of sensors, such as acoustic emission, vibration, temperature, and displacement, it covers all dimensions of gear machining information from cutting state to dimensional form, avoiding the one-sidedness of single-sensor monitoring. The adaptive preprocessing module dynamically reduces noise and performs spatiotemporal synchronous calibration to ensure data quality; the spatiotemporal feature fusion module extracts key temporal and spatial features from multi-source data, providing sufficient basis for quality prediction. This enables the system to capture quality change trends in real time during processing, provide early warning of quality risks, reduce rework and scrap losses, and improve production efficiency.

[0017] In terms of prediction accuracy and process adaptability, the improved deep belief network model accurately establishes the mapping relationship between process parameters and quality indicators through multi-layer RBM and attention mechanism. The prediction results can effectively reflect core quality parameters such as tooth surface roughness, tooth profile accuracy, and surface hardness. The introduction of transfer learning strategy enables the model to quickly adapt to new materials and working conditions without extensive retraining, solving the problem of poor generalization ability of traditional models. The online material property monitoring unit corrects the impact of material fluctuations on prediction in real time, further improving prediction accuracy and ensuring stable prediction of the machining quality of marine engineering gears of different materials and specifications.

[0018] In terms of decision-making practicality and system flexibility, the real-time decision feedback module can not only trigger early warnings or process adjustments based on prediction results, but also locate key influencing factors of quality defects through causal graph models, providing a clear direction for process optimization and avoiding the blindness of traditional adjustments. The edge-cloud collaborative architecture separates real-time prediction from offline model optimization. The edge rapidly processes data to meet real-time requirements, while the cloud continuously updates the model to ensure long-term adaptability. Dynamic data compression strategies reduce data transmission and storage pressure, balancing prediction accuracy and system operating efficiency. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the real-time quality prediction system for marine engineering gear processing based on multi-sensor fusion proposed in this invention. Figure 2 A comparison chart of the quality index errors of different prediction methods; Figure 3 This diagram illustrates the impact of process parameter adjustments on tooth surface roughness. Figure 4 A comparison chart showing the convergence speed of model transfer learning. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, 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 indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0023] Reference Figures 1 to 4 A real-time quality prediction system for marine engineering gear processing based on multi-sensor fusion, comprising the following modules: The multi-dimensional sensing data acquisition module employs a distributed sensor array, including 12-20 channels of acoustic emission sensors with a sampling rate of 1-3MHz, sensitivity of 75-85dB, and frequency range of 80kHz-1.2MHz; 6-10 channels of triaxial accelerometers with a range of ±40-±60g, resolution of 14-18 bits, and sampling rate of 40-60kHz; an infrared thermal imager with a resolution of 320×240-640×512, a temperature measurement range of -30℃-350℃, and a frame rate of 25-35Hz; and a laser displacement sensor with a measurement range of 0-100mm, accuracy of ±1-±3μm, and sampling rate of 0.5-2kHz. The sensor is connected to the data aggregation unit via an industrial Ethernet network. The industrial Ethernet network has a transmission rate of 0.5-2Gbps and a latency of ≤2ms. It collects acoustic emission signals, vibration acceleration, tooth surface temperature field distribution, and tooth profile dimensional deviation data during gear machining. Simultaneously, it records process parameters such as spindle speed, feed rate, and depth of cut. The spindle speed is 500-3000rpm, the feed rate is 0.05-0.3mm / r, and the depth of cut is 0.5-5mm. The data is packaged and stored as structured data frames with timestamps, and the timestamp accuracy is 1-5μs. The adaptive data preprocessing module performs multi-step cleaning and enhancement on the acquired raw data: A wavelet threshold denoising algorithm is used to remove Gaussian white noise from acoustic emission and vibration signals. This algorithm uses a db3-db5 wavelet basis, decomposes the data into 4-6 layers, and adaptively adjusts the threshold to 1.2-1.8 times the noise standard deviation. Kalman filtering is used to smooth the tooth profile size data of the laser displacement sensor. The state equation dimension of the Kalman filter is 4-8, the observation equation dimension is 2-4, and the process noise variance is 0.005-0.02. The measurement noise variance is 0.05-0.2. Non-uniformity correction and temperature calibration are performed on the infrared thermograms. Non-uniformity correction is based on the two-point correction method, and the correction coefficient is updated in real time every 0.5-2 minutes. The temperature calibration error is ≤±1.5℃. For missing data, a modified Lagrange interpolation method is used to complete the data. The proportion of missing data is <5%, and the interpolation order of the modified Lagrange interpolation method is 2-4. Outliers are replaced by the median value in a sliding window. Outliers are identified by the 3σ criterion, and the sliding window size is 30-70 sampling points. The spatiotemporal feature fusion module extracts the temporal and spatial features of multi-source data: For acoustic emission signals, it calculates kurtosis, peak factor, and energy entropy, with kurtosis ranging from 2 to 10, peak factor from 4 to 18, and energy entropy from 0.5 to 4; it performs Fourier transform on vibration signals with a frequency resolution of 0.5-2Hz, extracting 2-5 characteristic frequency amplitudes within the 80Hz-6kHz frequency band; for temperature field data, it extracts the highest tooth surface temperature, temperature gradient, and temperature uniformity coefficient, with a temperature gradient ≤8℃ / mm and a temperature uniformity coefficient ≥0.7; for tooth profile dimension data, it calculates tooth thickness deviation, cumulative tooth pitch error, and tooth profile error, with a tooth thickness deviation of ±0.02mm, cumulative tooth pitch error ≤0.03mm, and tooth profile error ≤0.02mm; and it dynamically assigns feature weights through an attention mechanism, with weights ranging from 0 to 1 and a total weight of 1, constructing a feature matrix containing 28-36 dimensional fused feature vectors. The quality prediction model module employs an improved Deep Belief Network (DBN) architecture: the input layer receives 28-36 dimensional fused feature vectors, the hidden layers contain 2-4 layers of Restricted Boltzmann Machines (RBMs) with 48-80, 24-40, and 12-20 neurons respectively, and the activation function is LeakyReLU with a negative slope of 0.005-0.02; the output layer consists of 3 fully connected neurons, outputting predicted values ​​for tooth surface roughness, tooth profile accuracy, and surface hardness, respectively, where tooth surface roughness is Ra0.5-4μm and tooth profile accuracy is IT4-I. T9 surface hardness is 180-400 HB; the model training uses the Adam optimizer with an initial learning rate of 0.0005-0.002, weight decay of 0.00005-0.0002, β1=0.85-0.95, β2=0.99-0.9999, and the loss function is root mean square error (RMSE). During training, the validation set accuracy is calculated every 30-70 epochs. When the accuracy improvement is <0.8% for 8-12 consecutive epochs, a learning rate decay strategy is adopted with a decay coefficient of 0.4-0.6. The real-time decision feedback module compares the predicted results with preset quality thresholds, which are: tooth surface roughness Ra≤2μm, tooth profile accuracy≥IT7, and surface hardness230-320HB. When a single indicator exceeds the threshold by less than 10%, an early warning signal is generated and the deviation value is displayed. When it exceeds the threshold by more than 10%, a process parameter adjustment command is triggered, and the feed rate and spindle speed are dynamically corrected through a PID controller. The feed rate adjustment step is 0.005-0.02mm / r, and the spindle speed adjustment step is 50-150rpm. At the same time, a quality-parameter association database is built to store the feature data, predicted results, and actual detection values ​​for each processing operation. The sampling frequency of the actual detection value is once per workpiece, supporting historical data backtracking and process optimization analysis. The system's collaborative control module adopts an edge-cloud collaborative architecture: the edge is an industrial-grade edge computing gateway with a quad-core ARM Cortex-A53 CPU and 2-8GB of memory, responsible for real-time data preprocessing and prediction, with a single sample processing time of <80ms; the cloud platform is equipped with a GPU cluster, including 2-6×NVIDIA A100, responsible for offline model training and updates; the edge and cloud communicate via a 5G private network with a transmission rate of 30-80Mbps and a latency of ≤30ms. The edge uploads feature data and prediction results every 0.5-2 hours, and the cloud pushes model update parameters every 12-48 hours. The model update adopts an incremental update strategy, with a data transmission volume of <15MB.

[0024] This invention also includes a multi-sensor spatiotemporal synchronization calibration unit, which achieves sensor time synchronization (synchronization error ≤ 8μs) through a hardware trigger signal (rising edge trigger, pulse width 5-15μs); spatial calibration employs a three-dimensional coordinate transformation algorithm, through... Transform the measurement points in each sensor coordinate system to the machine tool coordinate system, where These are the three-dimensional coordinates (unit: mm) in the machine tool coordinate system. R represents the three-dimensional coordinates (in mm) in the sensor coordinate system, R is a 3×3 rotation matrix (obtained by the three-point calibration method, with a rotation angle accuracy of ±0.2°), and T is a 3×1 translation vector (in mm, with an error ≤0.08 mm). The synchronous calibration results are displayed in real time on the system monitoring interface, and a calibration prompt is issued when the deviation exceeds the limit.

[0025] This invention also includes a processing condition adaptive adjustment submodule, which dynamically optimizes process parameters based on the matching degree between the prediction results and the current working conditions: calculating the working condition matching degree. , where n is the number of process parameters (n=3). The parameter weights are as follows (total weight is 1, spindle speed weight is 0.3-0.5, feed rate weight is 0.2-0.4, and depth of cut weight is 0.2-0.4). The current parameter value. For reference parameter values, k is the attenuation coefficient (k=0.3-0.7); when M<0.6-0.8, the parameter optimization algorithm is activated to search for the optimal parameter combination within the preset parameter range (spindle speed 500-3000rpm, feed rate 0.05-0.3mm / r, depth of cut 0.5-5mm) so that the predicted quality indicators meet the standards and the machining efficiency is improved by ≥8-12%.

[0026] In this invention, the adaptive data preprocessing module employs a variational mode decomposition (VMD) algorithm to decompose the acoustic emission signal, taking into account its non-stationary characteristics: through... Solve for the optimal modal components, where It is a set of K modal components (K=2-6). For the set of center frequencies of each mode, is the Dirac function, * is the convolution operation, t is the time variable (in seconds), and j is the imaginary unit; after decomposition, the top 2-4 modal components with the highest energy percentage are selected to reconstruct the signal, resulting in a signal-to-noise ratio improvement of ≥15-25dB.

[0027] In this invention, the spatiotemporal feature fusion module introduces a feature mutual information weighting mechanism to calculate the mutual information between any two features. ,in , These are two different features. Let be the joint probability density function. , The marginal probability density function is used; the mutual information value is normalized (0-1) as the feature association strength. Features with an association strength <0.2 are assigned 0.4-0.6 times the weight, and features with an association strength ≥0.2 are assigned 0.9-1.1 times the weight, thereby enhancing the targeting and effectiveness of feature fusion.

[0028] In this invention, the quality prediction model module employs a transfer learning strategy to optimize the model's adaptability to new material gear processing scenarios: the model trained based on low-carbon steel gear data is used as the source model, through... Construct the migration loss function, where For the loss in the target domain (high-strength alloy steel gear), The regularization coefficient is (0.005-0.2). For the source model weights, Set the target model weights; freeze the parameters of the first 1-3 layers of the source model RBM, and train only the last layer of RBM and the output layer. In the new scenario, the model convergence speed is improved by ≥40-60%, and the prediction error is reduced by ≥10-20%.

[0029] In this invention, the real-time decision feedback module constructs a quality traceability causal graph model and calculates the contribution of each process parameter to quality defects using a Bayesian network. ,in Let i be the i-th process parameter (i=1,2,3). Let j be the quality defect of type j (j=1,2,3). This represents the conditional probability of a defect occurring when parameters are abnormal. This represents the prior probability of parameter anomalies. The marginal probability of a defect occurring; parameters with a contribution greater than 0.3-0.5 are marked as key influencing factors and adjustment schemes are given priority in the decision recommendations.

[0030] In this invention, the system collaborative control module employs a dynamic data compression and transmission strategy to compress the feature data uploaded from the edge: through... To achieve compressed sensing, in which This is the compressed data (dimensions 6-10). The random measurement matrix is ​​6×32-10×36 (following a Gaussian distribution, mean 0, variance 1 / 6-1 / 10), and X is the original feature vector (dimension 28-36); the compression ratio is 3:1-5:1, and the reconstruction error is ≤6%, which reduces the pressure of cloud data storage and transmission while ensuring prediction accuracy.

[0031] This invention also includes an online monitoring unit for gear material properties, integrating an eddy current sensor (measurement frequency 0.8-1.2MHz, resolution 0.05-0.2μm) and an ultrasonic flaw detector (probe frequency 4-6MHz, detection depth 0-25mm) to collect real-time data on the hardness distribution (measurement interval 0.5-2mm) and internal defects (sensitivity ≥φ0.3-φ0.7mm flat-bottomed holes) of the gear blank; through... The input to the quality prediction model is modified, where H is the modified material hardness (in HB). The initial hardness (in HB) is given by E, the ultrasonic echo energy (in mV) is given by E, and α and β are correction coefficients (α=0.7-0.9, β=0.1-0.3, calibrated experimentally) to reduce the prediction error caused by material fluctuations by ≥15-25%.

[0032] In this invention, the quality prediction model module incorporates an online model evaluation and update triggering mechanism: after processing 8-12 gears, the average absolute percentage error between the predicted value and the actual detected value is calculated. Where n = 8 - 12, This is the actual measured value. The predicted value is used. When MAPE > 4-6%, incremental training in the cloud is automatically triggered. The model parameters are updated using 40-60 newly collected samples. During the update process, the system continues to operate normally (switching to the backup model) to ensure the continuity and accuracy of the prediction.

[0033] The following two examples further illustrate the specific implementation of this system: Example 1: Prediction of machining quality for low-speed, heavy-load gears in marine engineering (Material: high-strength alloy steel) I. System Deployment and Parameter Configuration Multi-dimensional sensor data acquisition module setup: A 16-channel acoustic emission sensor (SAEU2S, sampling rate 2MHz, sensitivity 80dB, frequency range 100kHz-1MHz) is distributed on the CNC gear hobbing machine (model YK3180) and installed on the machine tool spindle box, worktable, and tool holder; an 8-channel triaxial accelerometer (ADXL355, range ±50g, resolution 16-bit, sampling rate 50kHz) is fixed on the tool holder and workpiece fixture; an infrared thermal imager (FLIRA655sc, resolution 640×512, temperature range -20℃-300℃, frame rate 30Hz) is mounted above the machine tool with the lens aimed at the gear surface machining area; a laser displacement sensor (KeyenceLK-G80, measurement range 0-50mm, accuracy ±2μm, sampling rate 1kHz) is mounted on the machine tool slide and vertically aligned with the gear profile surface. The sensor is connected to the Siemens S7-1500 data aggregation unit via Profinet industrial Ethernet (transmission rate 1Gbps, delay ≤1ms) to synchronously record spindle speed 1200-2500rpm, feed rate 0.1-0.25mm / r, and depth of cut 1-3mm. The data is packaged into JSON format data frames with 1μs precision timestamps.

[0034] Adaptive data preprocessing module parameter debugging: The acoustic emission signal adopts the db4 wavelet basis, with 5 decomposition layers, and the threshold is set to 1.5 times the noise standard deviation; the vibration signal Kalman filter state equation has a dimension of 6 (position, velocity, acceleration), the observation equation has a dimension of 3, the process noise variance is 0.01, and the measurement noise variance is 0.1; the infrared thermogram non-uniformity correction period is 1 min, and the temperature calibration error is controlled within ±1℃; missing data (proportion <3%) is completed using third-order Lagrange interpolation, and outliers (3σ criterion: such as vibration acceleration > 8g) are replaced with the median of the sliding window of 50 sampling points. The preprocessed data is stored on a 128GB SSD at the edge.

[0035] Quality prediction model training: 5000 sets of high-strength alloy steel gear machining samples were collected (tooth surface roughness Ra0.8-3.2μm, tooth profile accuracy IT5-IT8, surface hardness 200-350HB), and the training and validation sets were divided in an 8:2 ratio. The improved DBN model received 32-dimensional fused features in the input layer, and the hidden layers consisted of 3 RBM neurons with 64, 32, and 16 neurons respectively, using the LeakyReLU activation function with a negative slope of 0.01. The Adam optimizer was used (initial learning rate 0.001, weight decay 0.0001, β1=0.9, β2=0.999), with the RMSE loss function. Validation accuracy was calculated every 50 epochs. If the accuracy improvement was <0.5% for 10 consecutive epochs, the learning rate was decayed by a factor of 0.5. The final model validation accuracy was ≥96%.

[0036] II. Real-time Quality Prediction and Decision Execution Spatiotemporal feature fusion and prediction: Kurtosis 3-8, peak factor 5-15, and energy entropy 1-3 are calculated for acoustic emission signals; Fourier transform of vibration signals (frequency resolution 1Hz) extracts characteristic frequency amplitudes at 100Hz, 500Hz, and 1kHz; temperature field data extracts the highest tooth surface temperature (80℃-150℃), temperature gradient ≤5℃ / mm, and temperature uniformity coefficient ≥0.8; tooth profile dimension data calculates tooth thickness deviation ±0.01mm, cumulative tooth pitch error ≤0.02mm, and tooth profile error ≤0.015mm. Feature weights are assigned using an attention mechanism (acoustic emission signal weight 0.3, vibration weight 0.25, temperature weight 0.2, dimension weight 0.25), constructing a 32-dimensional feature matrix input model. At the edge (NVIDIA Jetson AGXXavier, quad-core ARM Cortex-A57 CPU, 8GB memory), single-sample processing time is <50ms, outputting quality prediction values.

[0037] Real-time decision-making and process adjustment: The preset quality thresholds are tooth surface roughness Ra≤1.6μm, tooth profile accuracy≥IT6, and surface hardness250-300HB. When the predicted tooth surface roughness Ra=1.8μm (exceeding the threshold by 12.5%), the system triggers the PID controller to reduce the feed rate from 0.25mm / r to 0.22mm / r (adjustment step size 0.01mm / r) and increase the spindle speed from 2000rpm to 2200rpm (adjustment step size 100rpm). At the same time, the contribution of process parameters is calculated through a Bayesian network, with the spindle speed contributing 0.45 (a key factor), and the speed adjustment suggestion is displayed first on the decision interface.

[0038] Multi-sensor spatiotemporal synchronization calibration: Sensor time synchronization (error ≤ 5μs) is achieved through hardware trigger signals (rising edge trigger, pulse width 10μs); spatial calibration adopts the three-point calibration method, selecting three standard coordinate points on the workpiece surface ((0,0,0), (100,0,0), (0,100,0)), obtaining the rotation matrix R (rotation angle accuracy ±0.1°) and translation vector T (error ≤ 0.05mm), and then using the formula... ( For machine tool coordinate system coordinates, The calibration results are displayed in real time on the WinCC monitoring interface after the transformation is completed (to the sensor coordinate system coordinates). An audible and visual alert is issued when the deviation exceeds 0.1mm.

[0039] III. Effectiveness Verification Data Table 1

[0040] Table 1 verifies the advantages of this embodiment in the machining of high-strength alloy steel gears. Traditional post-processing inspection requires offline measurement after machining, which lags by 2-3 hours and cannot intervene in advance, resulting in a rework rate of up to 12%. This system, through multi-sensor fusion and real-time prediction, can simultaneously acquire predicted values ​​of three types of quality indicators, keeping the error within a small range and avoiding lag. The process adjustment response time is only 5 seconds, far faster than the traditional manual adjustment of more than 30 minutes, which can promptly correct parameter deviations and reduce the risk of quality defects. At the same time, spatiotemporal synchronous calibration ensures the consistency of multi-source data, and Bayesian network locates key parameters, making adjustments more targeted and significantly improving machining efficiency and quality stability.

[0041] Example 2: Prediction of Machining Quality for High-Speed ​​Lightly Loaded Gears in Marine Engineering (Material: Low Carbon Steel) I. System Adaptation and Model Optimization Multi-dimensional sensor data acquisition module adjustments: For the machining characteristics of low-carbon steel gears (model M20×2), sensor parameters were adjusted: the number of acoustic emission sensor channels was reduced to 12 (sampling rate 1.5MHz, sensitivity 78dB), the accelerometer had 6 channels (range ±40g, sampling rate 40kHz), the infrared thermal imager's temperature measurement range was adjusted to -10℃-250℃, and the laser displacement sensor's sampling rate was reduced to 0.8kHz. The data aggregation unit uses a Mitsubishi Q-series PLC, transmitting data via EtherNet / IP industrial Ethernet (transmission rate 0.8Gbps, delay ≤1.5ms), synchronously recording spindle speed 2000-3000rpm, feed rate 0.08-0.2mm / r, and depth of cut 0.5-2mm.

[0042] Transfer learning model training: The model trained on the low-carbon steel gear processing data in Example 1 is used as the source model, and the target domain is the current low-carbon steel gear processing scenario (sample size 800 groups). A transfer loss function is constructed. ( For the target domain loss, =0.05 regularization coefficient, Source model weights (Target model weights), freeze the parameters of the first two RBM layers of the source model, and train only the last RBM layer and the output layer. The training cycle is shortened from 30 days to 10 days, and the model convergence speed is significantly improved.

[0043] Dynamic data compression and cloud collaboration: The edge device (Advantech UNO-2484G, quad-core Celeron J4125 CPU, 4GB RAM) uses a compressed sensing algorithm on the 32-dimensional feature vector, through the formula... compression( For 8-dimensional compressed data, The model uses an 8×32 Gaussian random measurement matrix (mean 0, variance 1 / 8), with a compression ratio of 4:1 and a reconstruction error ≤5%. Every hour, the edge device uploads compressed data to the Alibaba Cloud GPU cluster (2×NVIDIA A100) via a 5G private network (transmission rate 50Mbps, latency ≤20ms). The cloud pushes incremental parameter updates every 24 hours (transmission data volume <10MB), and the edge device automatically updates the model upon receiving these updates.

[0044] II. Material Property Monitoring and Quality Control Online monitoring of gear material properties: An integrated eddy current sensor (model KAMANKD-2306, measurement frequency 1MHz, resolution 0.1μm) and ultrasonic flaw detector (model OlympusEPOCH650, probe frequency 5MHz, detection depth 0-20mm) are installed at the machine tool feed end to collect real-time data on the hardness distribution (measurement interval 1mm) and internal defects (sensitivity ≥ φ0.5mm flat-bottomed hole) of low-carbon steel billets. This data is obtained through formula... Corrected model input (H-corrected hardness, Initial hardness, E-wave energy, =0.8, =0.2), the prediction error caused by material fluctuations was significantly reduced.

[0045] Online model evaluation and update: For every 10 gears processed, calculate the mean absolute percentage error between the predicted value and the actual detected value. (n=10, Actual value (Predicted value). When MAPE > 5%, incremental training in the cloud is automatically triggered, using 50 newly collected samples to update the model. During the update process, the model is switched to a backup model (accuracy deviation from the main model < 2%) to ensure uninterrupted prediction.

[0046] III. Effectiveness Verification Data Table 2

[0047] Table 2 highlights the adaptation advantages of this embodiment in low-carbon steel gear processing scenarios. Traditional non-transfer learning models require 30 days to converge, while this system's transfer learning model, through parameter reuse, can adapt to new scenarios in just 10 days, significantly shortening the deployment cycle. Online monitoring of material properties corrects for the impact of material fluctuations, reducing prediction errors from "large" to "small," thus solving the prediction bias problem caused by uneven hardness in low-carbon steel. Dynamic data compression reduces the transmission volume from 100MB / hour to 25MB / hour, reducing bandwidth usage. Switching to a backup model during model updates results in an interruption time of only 10 minutes, far superior to the more than one hour of traditional systems, ensuring processing continuity. Overall, the system, through transfer learning, material monitoring, and dynamic compression, achieves rapid adaptation and efficient management across multiple scenarios, meeting the diverse processing needs of marine engineering gears.

[0048] Reference Figure 2 The figure clearly highlights the core advantage of this invention's system in terms of quality prediction accuracy. Traditional post-processing inspection relies on offline measurements after machining, which are susceptible to human error and environmental temperature and humidity interference, generally resulting in errors exceeding 14%, and cannot intervene in quality defects in advance. Single-sensor prediction relies only on single-dimensional data such as vibration or temperature, which cannot cover the entire process information such as cutting state and dimensional shape, and the error still remains at 8%-13%, which is difficult to meet the high precision requirements of marine engineering gears. This invention collects full-dimensional machining data through multi-sensor fusion, filters noise and corrects deviations through adaptive preprocessing, and then establishes a precise mapping relationship between features and quality indicators through an improved deep belief network model. Ultimately, the error of each quality indicator is controlled within 5.2%, far below the industrial-grade accuracy threshold. Among them, the tooth profile accuracy error is only 3.5%, which is due to the key features such as vibration frequency amplitude and tooth surface temperature gradient extracted by the spatiotemporal feature fusion module, as well as the attention mechanism's weight enhancement of highly correlated features, providing a reliable data basis for subsequent process parameter adjustments.

[0049] Reference Figure 3This figure fully demonstrates the effectiveness of the real-time decision feedback module in adjusting the process. In marine engineering gear machining, spindle speed and feed rate are the core process parameters affecting tooth surface roughness: when the spindle speed is too low and the feed rate is too high, the cutting force will increase significantly, easily leading to obvious tool marks on the tooth surface and an increase in the Ra value; while when the spindle speed is too high and the feed rate is too low, although the Ra value can be reduced, the machining time will be greatly extended, reducing production efficiency. This invention monitors the predicted value of tooth surface roughness in real time through a quality prediction model. When the predicted value exceeds the quality threshold by more than 10%, the PID controller is automatically triggered to adjust the parameters according to the preset adjustment step size; at the same time, the contribution of each parameter to the quality defect is calculated by combining a Bayesian network, and parameters with high contribution are adjusted first. After adjustment, the Ra value at the same spindle speed is reduced by an average of 46%, and at the parameter combination of 2200 rpm and 0.18 mm / r, Ra = 0.9 μm, which meets the quality threshold requirements and maintains high machining efficiency.

[0050] Reference Figure 4 This figure verifies the effect of the transfer learning strategy of this invention on improving the model's generalization ability. In the gear processing of marine engineering, different materials often need to be switched. Traditional quality prediction models need to be trained from scratch for new materials, but the number of samples for new materials is small and the collection period is long, resulting in slow model convergence: the accuracy is only 89.5% after 80 epochs of training, and it takes 88 epochs to reach the target accuracy of 95%, which cannot meet the production needs of multiple varieties and small batches. This invention adopts a transfer learning strategy, using a pre-trained low-carbon steel gear model as the "source model". By leveraging the general processing features already learned by the source model, only some model parameters are adjusted for the characteristics of the new material. The parameters of the first two layers of the restricted Boltzmann machine in the source model are frozen, and only the last layer of the restricted Boltzmann machine and the output layer are trained. This method significantly reduces the sample requirements for new scenarios. The accuracy reaches 81.5% after 20 epochs of training, and the target accuracy of 95% is reached after 65 epochs. The convergence speed is improved by 26% compared with traditional training, and the final accuracy reaches 97.8%. This strategy effectively shortens the deployment cycle of the model in new material scenarios and meets the adaptation requirements of multi-material processing of gears in marine engineering.

[0051] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A real-time quality prediction system for marine engineering gear machining using multi-sensor fusion, characterized in that, Includes the following modules: The multi-dimensional sensing data acquisition module adopts a distributed sensor array, including acoustic emission sensors, triaxial acceleration sensors, infrared thermal imagers, and laser displacement sensors. The sensors are connected to the data aggregation unit via industrial Ethernet to collect data during gear processing. The data is packaged and stored as structured data frames according to timestamps. The adaptive data preprocessing module removes Gaussian white noise from the acquired data using a wavelet threshold denoising algorithm; smooths the tooth profile size data of the laser displacement sensor using Kalman filtering; performs non-uniformity correction and temperature calibration on the infrared thermal image; and completes missing data using an improved Lagrange interpolation method, replacing outliers with the median value in a sliding window. The spatiotemporal feature fusion module calculates kurtosis, peak factor, and energy entropy for acoustic emission signals; performs Fourier transform on vibration signals to extract characteristic frequency amplitudes; extracts the highest tooth surface temperature, temperature gradient, and temperature uniformity coefficient from temperature field data; calculates tooth thickness deviation, cumulative tooth pitch error, and tooth profile error from tooth profile size data; and dynamically allocates feature weights through an attention mechanism to construct a feature matrix. The quality prediction model module receives 28-36 dimensional fused feature vectors as input, and the hidden layers contain 2-4 layers of restricted Boltzmann machines with 48-80, 24-40, and 12-20 neurons respectively. The output layer outputs the predicted values ​​of tooth surface roughness, tooth profile accuracy, and surface hardness respectively. The model is trained using the Adam optimizer with the root mean square error as the loss function, and the validation set accuracy is calculated every 30-70 epochs during training. The real-time decision feedback module compares the predicted results with the preset quality thresholds: when a single indicator exceeds the threshold by less than 10%, an early warning signal is generated and the deviation value is displayed; when it exceeds the threshold by more than 10%, a process parameter adjustment command is triggered, and the feed rate and spindle speed are dynamically corrected through the PID controller. The system's collaborative control module, at the edge, is responsible for data preprocessing and prediction; The cloud platform is responsible for offline model training and updates; The edge and cloud communicate via a 5G private network. The edge device uploads feature data and prediction results every 0.5-2 hours, while the cloud pushes model update parameters every 12-48 hours.

2. The multi-sensor fusion real-time quality prediction system for marine engineering gear processing according to claim 1, characterized in that, It also includes a multi-sensor spatiotemporal synchronization calibration unit, which synchronizes sensor time via hardware trigger signals; spatial calibration employs a three-dimensional coordinate transformation algorithm, through... Transform the measurement points in each sensor coordinate system to the machine tool coordinate system, where Three-dimensional coordinates in the machine tool coordinate system. R represents the three-dimensional coordinates in the sensor coordinate system, R is a 3×3 rotation matrix, and T is a 3×1 translation vector; the synchronous calibration results are displayed in real time on the system monitoring interface, and a calibration prompt is issued when the deviation exceeds the limit.

3. The multi-sensor fusion real-time quality prediction system for marine engineering gear processing according to claim 1, characterized in that, It also includes a submodule for adaptive adjustment of processing conditions, which dynamically optimizes process parameters based on the matching degree between the prediction results and the current conditions: calculating the matching degree of the processing conditions. Where n is the number of process parameters, For parameter weights, The current parameter value. Here is the reference parameter value, and k is the attenuation coefficient; When M < 0.6-0.8, the parameter optimization algorithm is activated to search for the optimal parameter combination within the preset parameter range.

4. The multi-sensor fusion real-time quality prediction system for marine engineering gear processing according to claim 1, characterized in that, In the adaptive data preprocessing module, a variational mode decomposition algorithm is used to decompose the acoustic emission signal, taking into account its non-stationary characteristics: through... Solve for the optimal modal components, where For a set of K modal components, For the set of center frequencies of each mode, is the Dirac function, * is the convolution operation, t is the time variable, and j is the imaginary unit; after decomposition, the modal components with the highest energy percentage (2-4) are selected to reconstruct the signal.

5. The multi-sensor fusion real-time quality prediction system for marine engineering gear machining according to claim 1, characterized in that, The spatiotemporal feature fusion module introduces a feature mutual information weighting mechanism to calculate the mutual information between any two features. ,in , These are two different features. Let be the joint probability density function. , The marginal probability density function is used; the mutual information value is normalized as the feature association strength. Features with an association strength < 0.2 are assigned a weight of 0.4-0.6 times, and features with an association strength ≥ 0.2 are assigned a weight of 0.9-1.1 times, thereby enhancing the targeting and effectiveness of feature fusion.

6. The multi-sensor fusion real-time quality prediction system for marine engineering gear machining according to claim 1, characterized in that, In the quality prediction model module, a transfer learning strategy is used to optimize the model's adaptability to the new material gear processing scenario: the model trained based on low-carbon steel gear data is used as the source model, through... Construct the migration loss function, where For the target domain loss, The regularization coefficient is . For the source model weights, Set the weights for the target model; freeze the parameters of the first 1-3 layers of the source model's RBM, and train only the last RBM layer and the output layer.

7. The multi-sensor fusion real-time quality prediction system for marine engineering gear processing according to claim 1, characterized in that, In the real-time decision feedback module, a quality traceability cause-effect graph model is constructed, and a Bayesian network is used to calculate the contribution of each process parameter to quality defects. ,in For the i-th process parameter, For the j-th type of quality defect, This represents the conditional probability of a defect occurring when parameters are abnormal. This represents the prior probability of parameter anomalies. The marginal probability of a defect occurring; parameters with a contribution greater than 0.3-0.5 are marked as key influencing factors and adjustment schemes are given priority in the decision recommendations.

8. The multi-sensor fusion real-time quality prediction system for marine engineering gear machining according to claim 1, characterized in that, In the system collaborative control module, a dynamic data compression and transmission strategy is used to compress the feature data uploaded from the edge: through... To achieve compressed sensing, in which For compressed data, The random measurement matrix is ​​6×32-10×36, and X is the original eigenvector; the compression ratio is 3:1-5:1, and the reconstruction error is ≤6%.

9. The multi-sensor fusion real-time quality prediction system for marine engineering gear machining according to claim 1, characterized in that, It also includes an online monitoring unit for gear material properties, integrating an eddy current sensor and an ultrasonic flaw detector to collect real-time data on the hardness distribution and internal defects of the gear blank; through The input to the quality prediction model is modified, where H is the modified material hardness. α represents the initial hardness, E represents the ultrasonic echo energy, and α and β are correction coefficients.

10. The multi-sensor fusion real-time quality prediction system for marine engineering gear machining according to claim 1, characterized in that, The quality prediction model module incorporates an online model evaluation and update trigger mechanism: after processing 8-12 gears, the average absolute percentage error between the predicted value and the actual detected value is calculated. Where n = 8 - 12, This is the actual measured value. The values ​​are predicted. When MAPE > 4-6%, incremental training in the cloud is automatically triggered, using 40-60 newly collected samples to update the model parameters, while maintaining normal system operation during the update process.

Images