A big data bamboo tableware manufacturing system based on production process monitoring

Abstract

The present invention provides a big data bamboo tableware manufacturing system based on production process monitoring, which relates to the field of intelligent detection systems. The system comprises an acquisition module, a prediction module, a preliminary diversion module, a disturbance response module, an optimization decision module and an abnormality self-detection module. Images are collected by a high-speed camera, an initial confidence score is output using a pre-trained model, a dynamic stability index is calculated in combination with a dynamic image sequence under disturbance response, and a rejection target is determined by an optimization decision module after the scores are integrated. During the optimization process, a Bayesian method is introduced to dynamically adjust multiple decision parameters and generate rejection action instructions in real time. The abnormality self-detection module monitors the rejection rate and score stability, and multi-level fault diagnosis and system protection control are realized by combining software and hardware signals. The present invention realizes high-precision recognition and dynamic rejection of bamboo tableware defects, thereby improving the automation level and production reliability.

Description

Technical Field

[0001] The present invention relates to the field of intelligent detection technology, and in particular to a big data bamboo tableware manufacturing system based on production process monitoring. Background Art

[0002] The manufacturing and quality inspection of bamboo tableware mostly uses manual inspection or primary automated inspection systems based on static image processing, which are mainly used to identify defects on the surface of tableware, such as cracks, burrs, deformation, etc.

[0003] However, as a naturally heterogeneous material, bamboo is prone to structural defects such as hidden cracks, internal cracks, delamination, mildew, and insect-eaten channels. Although these defects are not easily detected by surface images, they will seriously affect the mechanical strength and safety of the product. Existing static image-based detection methods are difficult to accurately identify the above-mentioned internal or dynamic response defects, resulting in the inability to effectively eliminate some unqualified products, which restricts the further improvement of product quality and the intelligent development of the manufacturing process. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a big data bamboo tableware manufacturing system based on production process monitoring, which improves the recognition accuracy of bamboo tableware defects through a two-stage scoring mechanism that integrates image recognition and disturbance response.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] A big data bamboo tableware manufacturing system based on production process monitoring, including a transmission line and a turning mechanism, and also includes:

[0007] An acquisition module, used to collect image data of tableware through high-speed cameras deployed on both sides of the turning mechanism;

[0008] A prediction module, configured to output an initial confidence score for each tableware based on the image data using a pre-trained prediction model;

[0009] A preliminary diversion module is used to set a first decision threshold based on the initial confidence score. If the initial confidence score is lower than the first decision threshold, the dish is judged as a high-probability defective product and directly enters the rejection zone. If the initial confidence score is equal to or higher than the threshold, the dish is transported to the disturbance section of the transmission line;

[0010] A disturbance response module is configured to apply a preset airflow disturbance to tableware entering the disturbance section, obtain a dynamic image sequence of the tableware under the airflow disturbance from at least one perspective, calculate the dynamic stability index of the tableware using optical flow analysis, fuse the initial image features of the tableware with the dynamic stability index, process them using an ensemble learning algorithm, and output a fusion confidence score;

[0011] An optimization decision module is used to set the second decision threshold and fuzzy decision interval parameters, determine the final rejection target based on the comparison result of the fusion confidence score and the fuzzy decision interval, obtain historical rejection data and real-time status data of the production line, and take minimizing the false rejection rate and ensuring that the defect detection rate meets the requirements as the optimization goal. Based on the historical rejection data and real-time status data of the production line, the first decision threshold, the second decision threshold and the fuzzy decision interval parameters are dynamically determined through the Bayesian optimization method, and the updated first decision threshold parameters are sent to the preliminary diversion module in real time. The second decision threshold and the fuzzy decision interval parameters are retained for self-use. Based on the fusion confidence score and the dynamically adjusted second decision threshold, the rejection mechanism action instructions are generated.

[0012] Furthermore, the abnormal self-check module is used to continuously monitor the tableware rejection rate within a set time window and set a dynamic safety threshold. The dynamic safety threshold is dynamically updated based on a historical rejection rate statistical model. When the real-time rejection rate continues to exceed the dynamic safety threshold for a predetermined period of time, the multi-level self-check program is automatically triggered.

[0013] Furthermore, the prediction model architecture adopts a deep residual network, whose main structure includes 4 residual block groups, each group consists of 2 residual blocks; two convolutional layers are set in each residual block and cross-layer identity mapping operations are performed, and the degradation problem of deep networks is solved by skip connections, which is used to extract surface defects and geometric deformation features of bamboo tableware; a multi-scale feature fusion structure is embedded in the front end of the residual network: horizontal connection channels are established at the output ends of the second and third residual block groups respectively, and the bottom high-resolution feature maps and high-level strong semantic feature maps are input into the feature pyramid network; the high-resolution feature maps are aligned through bilinear interpolation upsampling operations. The resolution of the layer feature map is increased and element-by-element fused with the underlying feature map of the corresponding layer. The fused multi-scale feature map is output to simultaneously capture local microscopic defects and overall morphological features of the tableware. A dilated convolution layer is embedded in the third residual block group, with a dilation rate of no less than 2 to expand the receptive field and enhance sensitivity to subtle fracture defects in the bamboo texture without reducing resolution. A self-attention mechanism is integrated at the output of the last residual block group: by calculating the correlation weight matrix between the spatial positions of the feature map, the feature weights of each region are dynamically adjusted to focus the model on areas with high defect risk.

[0014] The output layer of the prediction model uses a multi-label classification framework: the fully connected layer contains 512 neurons and receives feature vectors weighted by self-attention. The output layer has the same number of independent sub-output layers as the number of defect types. Each sub-output layer uses a sigmoid activation function to output the probability value of the corresponding defect. The probability of each defect is weighted and summed according to the preset defect risk weight to generate an initial confidence score in the range of [0,1]. The lower the score, the higher the defect risk. The prediction model training process adopts a supervised learning strategy: using a dataset containing labeled samples, the Adam optimizer is used to minimize the multi-task cross entropy loss function.

[0015] Furthermore, the dynamic stability index of the tableware is calculated by optical flow analysis, including:

[0016] Based on a continuous image sequence of tableware under airflow disturbance obtained by the disturbance response module, the image acquisition frame rate is not less than 200 frames per second; an optical flow tracking algorithm is used to track the displacement of at least five key geometric feature points preset on the surface of the tableware in the image sequence, wherein the key geometric feature points include the center point of the tableware, the point of maximum edge curvature, and the joint point, and motion trajectory data of each feature point is generated; three characteristic parameters are extracted based on the motion trajectory data: the first is the maximum offset amplitude, which is obtained by calculating the straight-line distance of all feature points from the initial position to the maximum displacement point during the airflow disturbance, and taking the maximum value of the distance values ​​of all feature points as the maximum offset amplitude; the second is the posture recovery time, which is calculated from the moment the airflow disturbance ends, and the time interval in milliseconds during which the displacement of all feature points remains within the range of ±0.5 mm from the initial position for 100 milliseconds; the third is the resonance intensity coefficient, which is obtained by performing spectral analysis on the motion trajectory of the feature points and extracting the ratio of the energy of the main frequency component in the amplitude spectrum to the total energy to obtain the resonance intensity coefficient; and finally, the maximum offset amplitude, posture recovery time, and resonance intensity coefficient are combined to form a dynamic stability index.

[0017] Furthermore, the decision parameters dynamically determined by the Bayesian optimization method include:

[0018] The objective function is defined as minimizing the rejection misjudgment rate, while setting a constraint that the defect detection rate is no less than a predetermined standard value. Based on the fusion confidence score distribution in historical rejection data, the frequency of multiple defect types, and the real-time transmission speed characteristics of the production line, a Gaussian process regression algorithm is used to construct a mapping relationship between decision parameters and optimization objectives. The decision parameters include a first decision threshold, a second decision threshold, and the width of the fuzzy decision interval.

[0019] Dynamic adjustment is achieved by iteratively executing the following process: periodically generating multiple sets of candidate parameter combinations within the range of decision parameter values, using the Gaussian process regression algorithm to predict the error rate and defect detection rate corresponding to each combination, screening candidate combinations that meet the defect detection rate constraint and have an error rate lower than the currently used parameters, and selecting combinations with lower error rates to update parameter settings; outputting the updated first decision threshold to the preliminary diversion module, and outputting the second decision threshold and the fuzzy decision interval width to the optimization decision module.

[0020] Furthermore, the rejection mechanism action instructions generated by the optimization decision module include multiple control parameters for driving the rejection execution structure, and the control parameters include: the rejection nozzle airflow injection angle dynamically determined according to the real-time posture data of the target tableware on the transmission line; the pulse air pressure intensity level set according to the degree of deviation between the fusion confidence score and the currently used second decision threshold, and at least three air pressure intensities are set, and the greater the degree of deviation, the higher the corresponding air pressure intensity level; the nozzle rejection trigger timing determined based on the production line transmission speed and the predicted position of the target tableware.

[0021] Furthermore, the multi-level self-checking procedure includes:

[0022] Conduct signal integrity and operational stability diagnosis on the high-speed camera, disturbance airflow nozzle array, and transmission line encoder that constitute the disturbance section, completing equipment layer self-test;

[0023] During the device-level self-test or when no anomalies are found during the device-level self-test, drift detection analysis is performed on the initial confidence score sequence output by the prediction module, and stability analysis is performed on the fused confidence score sequence output by the disturbance response module to complete the algorithm-level self-test;

[0024] When both the equipment-level self-inspection and the algorithm-level self-inspection confirm the existence of an abnormality and the cause of the abnormality has not been located, a graded alarm signal is generated and pushed to the central monitoring platform to trigger manual intervention, and a pause operation instruction is sent to the production line control system.

[0025] Furthermore, if no abnormality is found in the device layer self-test, the algorithm layer self-test is performed, including:

[0026] Comparing and analyzing the real-time initial confidence score sequence currently output by the prediction module with a pre-established reference sequence of initial confidence scores during a stable operation benchmark period, and determining that the performance of the prediction model has drifted when the difference in statistical characteristics of the two sequences exceeds a preset drift threshold;

[0027] The real-time fusion confidence score sequence currently output by the disturbance response module is compared with the pre-established fusion confidence score reference sequence during the stable operation benchmark period for fluctuation analysis. When the fluctuation of the real-time sequence exceeds the preset fluctuation threshold or the corresponding indicator of the reference sequence, it is determined that the stability of the fusion confidence score output has decreased;

[0028] Performing pattern recognition on at least one of the real-time initial confidence score sequence output by the prediction module and the real-time fused confidence score sequence output by the disturbance response module, and determining that an abnormal pattern exists in the score output when detecting the presence of multiple consecutive identical score values ​​in the sequence, abnormal amplitude mutations in the score values, or frequent occurrence of extremely high or extremely low score values ​​within a short time window;

[0029] Pre-calibrated tableware benchmark samples are regularly input into the prediction module and the disturbance response module; the initial confidence score output by the prediction module and the fusion confidence score output by the disturbance response module are respectively compared with the preset benchmark scores corresponding to the benchmark samples; when the score deviation exceeds the preset allowable range, the abnormal information is recorded and an algorithm performance abnormality signal is sent to the optimization decision module.

[0030] Furthermore, the device layer self-test procedure includes:

[0031] The device layer self-test program specifically performs the following detection operations: sending a test signal to the high-speed camera to analyze the timing continuity and pixel integrity of the image frame; collecting the air pressure sensor readings of the disturbance airflow nozzle array in real time and comparing them with the target air pressure intensity set by the optimization decision module; reading the transmission line encoder displacement pulse, synchronously correlating the tableware posture data captured by the disturbance response module to calculate the position offset; recording the triggering moment of the rejection instruction and the actual response time of the rejection nozzle feedback signal.

[0032] Furthermore, the abnormal self-check module jointly analyzes the abnormal signals detected by the equipment layer, the real-time rejection rate fluctuation data and the score stability results output by the algorithm layer self-check; when at least two detection abnormalities in the equipment layer are met at the same time, the real-time rejection rate exceeds the dynamic safety threshold, and the algorithm layer self-check confirms that the initial confidence score of the prediction module drifts or the stability of the fusion confidence score decreases, a downgrade instruction is sent to the optimization decision module and a protection mechanism is triggered: dynamically increase at least one of the first decision threshold and the second decision threshold, or suspend the execution of the current rejection action, or switch to a simplified decision strategy based on the historical average confidence score.

[0033] The above solution of the present invention includes at least the following beneficial effects:

[0034] The above solution of the present invention, by introducing a two-stage scoring mechanism that integrates image recognition and disturbance response, is conducive to improving the recognition accuracy of bamboo tableware defects and increasing the detection rate of internal or structural defects such as hidden cracks and delamination.

[0035] The Bayesian optimization algorithm is used to dynamically adjust the decision threshold and fuzzy interval, which helps to reduce the false rejection rate and adapt to the recognition requirements of different batches and working conditions;

[0036] By integrating the dynamic stability index of optical flow analysis, the recognition of structural anomalies is improved, and the stability of production lines is increased;

[0037] The abnormality self-check module combines the algorithm layer and the equipment layer monitoring mechanism to realize the real-time diagnosis and response of the system performance drift and operation abnormality in the present invention, ensuring the continuous and stable operation of the production line. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the module relationship provided by the present invention.

[0039] Figure 2 It is a schematic diagram of the production line structure in the present invention.

[0040] In the figure: 101, transmission line; 102, flipping mechanism; 103, high-speed camera; 104, rejection nozzle; 105, rectangular box; 106, disturbed airflow nozzle; 107, encoder. DETAILED DESCRIPTION

[0041] The following describes exemplary embodiments of the present disclosure in more detail with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0042] like Figures 1 to 2 As shown, an embodiment of the present invention provides a big data bamboo tableware manufacturing system based on production process monitoring, including a transmission line 101 and a turning mechanism 102, and further comprising:

[0043] An acquisition module, configured to collect image data of tableware using high-speed cameras 103 disposed on both sides of the turning mechanism 102;

[0044] A prediction module, configured to output an initial confidence score for each tableware based on the image data using a pre-trained prediction model;

[0045] A preliminary diversion module is used to set a first decision threshold based on the initial confidence score. If the initial confidence score is lower than the first decision threshold, the dish is determined to be a high-probability defective product and is directly sent to the rejection zone. If the initial confidence score is equal to or higher than the threshold, the dish is transported to the disturbance section of the transmission line 101.

[0046] A disturbance response module is configured to apply a preset airflow disturbance to tableware entering the disturbance section, obtain a dynamic image sequence of the tableware under the airflow disturbance from at least one perspective, calculate the dynamic stability index of the tableware using optical flow analysis, fuse the initial image features of the tableware with the dynamic stability index, process them using an ensemble learning algorithm, and output a fusion confidence score;

[0047] An optimization decision module is used to set a second decision threshold and fuzzy decision interval parameters, determine the final rejection target based on the comparison result of the fusion confidence score and the fuzzy decision interval, obtain historical rejection data and real-time status data of the production line, and take minimizing the false rejection rate and ensuring that the defect detection rate meets the requirements as the optimization goal. Based on the historical rejection data and real-time status data of the production line, the first decision threshold, the second decision threshold and the fuzzy decision interval parameters are dynamically determined by the Bayesian optimization method, and the updated first decision threshold parameters are sent to the preliminary diversion module in real time. The second decision threshold and the fuzzy decision interval parameters are retained for self-use. Based on the fusion confidence score and the dynamically adjusted second decision threshold, a rejection mechanism action instruction is generated;

[0048] The abnormality self-check module is used to continuously monitor the tableware rejection rate within a set time window and set a dynamic safety threshold. The dynamic safety threshold is dynamically updated based on a historical rejection rate statistical model. When the real-time rejection rate continues to exceed the dynamic safety threshold for a predetermined period of time, a multi-level self-check program is automatically triggered.

[0049] In an embodiment of the present invention, the tableware to be tested is placed at the left end of the transmission line 101 by a loading robot arm, and the transmission line 101 transports the tableware to be tested from left to right. High-speed cameras 103 are symmetrically deployed on both sides of the flipping mechanism 102, and the high-speed camera 103 is arranged above the disturbance section. The high-speed camera 103 is used to collect image data of the tableware.

[0050] In the prediction module, the pre-trained prediction model is trained based on 200,000 historically collected bamboo tableware image data that are annotated with defect types and location information. The prediction model adopts a deep residual network architecture and extracts the complex surface features of the tableware through a cross-layer identity mapping structure.

[0051] In the preliminary diversion module, the first decision threshold is set at 0.6. When the initial confidence score output by the prediction module is lower than 0.6, it is judged as a high-probability defective product, and the high-probability defective tableware is rejected through the rejection areas set on both sides of the turning mechanism 102; if the initial confidence score is equal to or higher than 0.6, the tableware is transported to a disturbance section with a length of 2 meters through the transmission line 101.

[0052] In the disturbance response module, an array of multiple disturbance airflow nozzles 106 is deployed above the disturbance section to apply a pulsed airflow disturbance with a duration of 0.5 seconds and a flow rate of 15 m / s to the tableware. The airflow direction is perpendicular to the surface of the transmission line 101 and tilted downward at an angle of 30 degrees. The high-speed camera 103 obtains a dynamic image sequence of the tableware under airflow disturbance at a frame rate of 200 frames per second. The dynamic stability index of the tableware is calculated by the optical flow analysis method, and the initial image features of the tableware are fused with the dynamic stability index. The fusion is processed by an integrated learning algorithm to output a fusion confidence score.

[0053] In the optimization decision module, the second decision threshold is set to 0.75, and the fuzzy decision interval parameter is [0.7, 0.8]. The final rejection target is determined based on the comparison result of the fusion confidence score and the fuzzy decision interval. The historical rejection data and real-time status data of the production line in the past 7 days are obtained. The optimization goal is to minimize the false rejection rate and ensure that the defect detection rate is not lower than the preset value (95% in this embodiment). Based on the historical rejection data and the real-time status data of the production line, the first decision threshold, the second decision threshold and the fuzzy decision interval parameters are dynamically determined by the Bayesian optimization method. The updated first decision threshold parameters are sent to the preliminary diversion module in real time. The second decision threshold and the fuzzy decision interval parameters are retained for self-use. Based on the fusion confidence score and the dynamically adjusted second decision threshold, the optimal rejection mechanism action instructions are generated.

[0054] In the abnormal self-inspection module, the tableware rejection rate is continuously monitored within 1 hour. When the real-time rejection rate continues to exceed the dynamic safety threshold for 10 minutes, the multi-level self-inspection program is automatically triggered. The dynamic safety threshold is determined based on the moving average of the rejection rate in the past 24 hours. The baseline threshold fluctuates by ±20% as the dynamic safety threshold, thereby constructing a statistical model to dynamically update the safety threshold.

[0055] It should be noted that in this embodiment, the first decision threshold is set to 0.6, the second decision threshold is set to 0.75, and the fuzzy decision interval is set to [0.7, 0.8]. These are only example values ​​and can be adjusted appropriately according to production line conditions in actual applications.

[0056] In this embodiment, the surface of the transmission line 101 carrying the bamboo tableware is meshed, so that the rejection mechanism can blow the bamboo tableware away from the transmission line 101 through the ejected airflow passing through the meshed surface, thereby completing the rejection action.

[0057] A rectangular box 105 is provided on the transmission line 101. The left and right vertical surfaces of the rectangular box 105 are through-shaped. The transmission line 101 is arranged inside the rectangular box 105. The high-speed camera 103 and the turbulent airflow nozzle 106 are mounted on the inner top surface of the rectangular box 105. The section of the transmission line 101 located inside the rectangular box 105 is the turbulent section.

[0058] In this embodiment, the rejection mechanism is a rectangular array of rejection nozzles 104 provided below the transmission line 101 , and the orthographic projection area of ​​the rejection nozzles 104 on the transmission line 101 is the rejection zone.

[0059] The pre-trained prediction model outputs an initial confidence score for each dish, including:

[0060] The prediction model architecture adopts a deep residual network, and its main structure includes 4 residual block groups, each group consists of 2 residual blocks; two convolutional layers are set in each residual block and cross-layer identity mapping operations are performed, and the degradation problem of deep networks is solved by skip connections, which is used to extract surface defects and geometric deformation features of bamboo tableware; a multi-scale feature fusion structure is embedded in the front end of the residual network: horizontal connection channels are established at the output ends of the second and third residual block groups respectively, and the bottom high-resolution feature maps and high-level strong semantic feature maps are input into the feature pyramid network; the high-level feature maps are aligned through bilinear interpolation upsampling operations. The resolution of the feature map is increased and element-by-element fused with the underlying feature map at the corresponding level. The fused multi-scale feature map is output to simultaneously capture local microscopic defects and overall morphological features of the tableware. A dilated convolutional layer is embedded in the third residual block group, with a dilation rate of no less than 2 to expand the receptive field and enhance sensitivity to subtle fracture defects in the bamboo texture without reducing resolution. A self-attention mechanism is integrated at the output of the last residual block group: by calculating the correlation weight matrix between spatial positions in the feature map, the feature weights of each region are dynamically adjusted to focus the model on areas with high defect risk.

[0061] The output layer of the prediction model uses a multi-label classification framework: the fully connected layer contains 512 neurons and receives feature vectors weighted by self-attention. The output layer has the same number of independent sub-output layers as the number of defect types. Each sub-output layer uses a sigmoid activation function to output the probability value of the corresponding defect. The probability of each defect is weighted and summed according to the preset defect risk weight to generate an initial confidence score in the range of [0,1]. The lower the score, the higher the defect risk. The prediction model training process adopts a supervised learning strategy: using a dataset containing labeled samples, the Adam optimizer is used to minimize the multi-task cross entropy loss function.

[0062] In this embodiment of the present invention, the prediction model uses a deep residual network (ResNet-50) architecture, connecting different layers via a cross-layer identity mapping structure to extract the complex surface features of tableware. The deep residual network front-end constructs a multi-scale feature pyramid. In this embodiment of the present invention, the multi-scale feature fusion structure exemplarily uses 16×16, 32×32, and 64×64 convolution kernels to capture local microscopic defects and overall morphology.

[0063] A hole convolution layer is embedded in the residual block of the deep residual network. In this embodiment, the hole ratio is set to 2 to enhance the sensitivity to bamboo defects. A self-attention mechanism is integrated at the end of the deep residual network to focus on defect-related areas by calculating the attention weights of different areas. The output layer of the prediction model adopts a multi-label classification framework, and presets defect types such as cracks, burrs, deformations, and thickness anomalies. The risk weight is set according to the degree of influence of each defect on the quality of tableware. In this embodiment, the crack risk weight is set to 0.4, the burr risk weight is set to 0.3, the deformation risk weight is set to 0.2, and the thickness anomaly is set to 0.1. In actual applications, it can be adjusted according to needs, and the probability of each defect is aggregated to generate an initial confidence score;

[0064] The specific training process of the prediction model is as follows:

[0065] (1) At least 200,000 real bamboo tableware images with manually annotated defect types and locations were uniformly resized to 224×224 pixels and randomly divided into a training set (160,000 images) and a validation set (40,000 images) with a ratio of 8:2;

[0066] (2) Data augmentation processing was performed on the training set images, including random rotation (0-15 degrees), random horizontal flipping, random brightness adjustment (±10%), and random cropping (90% of the original image size and then restored to 224×224 size), and pixel normalization was performed to the range of [0,1];

[0067] (3) The Adam optimizer was used for model training. The initial learning rate was set to 0.001. The standard parameters of the Adam optimizer were β1 = 0.9, β2 = 0.999, and epsilon = 1e-8. The warm-up learning rate warm-up strategy was adopted. The learning rate was gradually increased from 0.0001 to 0.001 in the first five epochs. The learning rate was reduced to 0.1 times the original value at the end of the 30th, 60th, and 90th epochs respectively.

[0068] (4) The training process is carried out for a maximum of 100 epochs, with a batch size of 64 per epoch. The prediction model calculates Precision, Recall, and F1-score on the validation set. If the validation set loss value does not improve for 10 consecutive epochs, the training is stopped.

[0069] (5) The model parameters corresponding to the highest F1-score of the validation set are used as the final prediction model parameters. The training and validation process uses the TensorFlow framework, which is well-known in the field of deep learning, and is completed on a server equipped with an NVIDIA Tesla V100 GPU.

[0070] The dynamic stability index of tableware is calculated by optical flow analysis, including:

[0071] Based on a continuous image sequence of tableware under airflow disturbance acquired by the disturbance response module, the image acquisition frame rate is not less than 200 frames per second. An optical flow tracking algorithm is used to track the displacement of at least five key geometric feature points preset on the surface of the tableware in the image sequence. The key geometric feature points include the center point of the tableware, the point of maximum edge curvature, and the joint point, and motion trajectory data of each feature point is generated. Three characteristic parameters are extracted from the motion trajectory data: the first is the maximum offset amplitude, which is obtained by calculating the straight-line distance of all feature points from the initial position to the point of maximum displacement during the airflow disturbance, and taking the maximum value of all feature point distance values ​​as the maximum offset amplitude; the second is the posture recovery time, which is calculated from the moment the airflow disturbance ends, and the time interval (in milliseconds) during which the displacement of all feature points remains within a range of ±0.5 mm from the initial position for 100 milliseconds; the third is the resonance intensity coefficient, which is obtained by performing spectral analysis on the motion trajectory of the feature points, extracting the ratio of the energy of the main frequency component in the amplitude spectrum to the total energy, multiplying the ratio by the main frequency value and dividing by 10 to obtain the resonance intensity coefficient; and finally, the maximum offset amplitude, posture recovery time, and resonance intensity coefficient are combined to form a dynamic stability index.

[0072] In an embodiment of the present invention, the Lucas-Kanade optical flow algorithm is used to perform optical flow field calculation on the acquired dynamic image sequence of tableware. The image acquisition frame rate is 200 frames per second. The displacement vectors and vibration frequency information of no less than five key geometric feature points preset on the tableware surface are extracted. In this embodiment, the displacement vectors and vibration frequency information of 50 key feature points on the tableware surface are extracted.

[0073] Based on the above calculation results, it is further clarified that:

[0074] The maximum offset amplitude is obtained by comparing the displacement vector lengths of all 50 key feature points and selecting the maximum displacement vector length;

[0075] The attitude recovery time is measured from the end of the airflow disturbance, and the time it takes for the displacement of all 50 key feature points to return to and remain within ±0.5 mm of the initial position for 100 milliseconds is determined;

[0076] The resonance intensity coefficient is obtained by performing spectral analysis on the motion trajectories of all 50 feature points, determining the main frequency component in the amplitude spectrum, multiplying the energy proportion of the main frequency by the main frequency value and dividing by 10. The step of dividing by 10 is an exemplary setting to facilitate the normalization of the resonance intensity coefficient value, and can be adjusted according to needs in actual applications.

[0077] Dynamically determining decision parameters through Bayesian optimization methods includes:

[0078] The objective function is defined as minimizing the rejection misjudgment rate, and the defect detection rate is set to be no less than a predetermined standard value as a constraint condition; based on the fusion confidence score distribution in the historical rejection data, the frequency of occurrence of multiple types of defects and the real-time transmission speed characteristics of the production line, the Gaussian process regression algorithm is used to construct the mapping relationship between the decision parameters and the optimization target. The decision parameters include the first decision threshold, the second decision threshold and the fuzzy decision interval width; dynamic adjustment is achieved by iteratively executing the following process: multiple groups of candidate parameter combinations are periodically generated within the range of decision parameter values, and the Gaussian process regression algorithm is used to predict the misjudgment rate and defect detection rate corresponding to each combination. The candidate combinations that meet the defect detection rate constraint and have an error rate lower than the currently used parameters are screened, and the combination with a lower error rate is selected to update the parameter settings; the updated first decision threshold is output to the preliminary diversion module, and the second decision threshold and the fuzzy decision interval width are output to the optimization decision module.

[0079] In this embodiment of the present invention, the process for dynamically determining decision parameters using a Bayesian optimization method is as follows: The optimization objective is defined as minimizing the false reject rate, with a constraint that the defect detection rate must be no less than 95%. The input data for the optimization process includes the statistical distribution of fused confidence scores from the past seven days of historical reject data, the frequency of various defect types in the historical data, and the current production line's real-time throughput (measured in the number of dishes processed per minute).

[0080] During the optimization process, the aforementioned historical data and real-time production line data are first used to construct a proxy model between the decision parameters (the first decision threshold, the second decision threshold, and the fuzzy decision interval width) and the false rejection rate and defect detection rate using the Gaussian process regression algorithm. Then, a parameter optimization iteration is performed every 30 minutes: multiple candidate parameter combination sets are periodically generated within the allowable value range of the decision parameters (in this embodiment, 10 candidate combinations are generated at a time). The mapping relationship constructed by the Gaussian process regression algorithm is used to predict the false rejection rate and defect detection rate for each candidate parameter combination. From these candidate parameter combinations, candidate parameters that both meet the defect detection rate constraint (no less than 95%) and have a predicted false rejection rate lower than the currently used parameter combination are selected. The parameter combination with the lowest false rejection rate is selected as the new decision parameter and updated to the system in real time. The updated first decision threshold is sent to the preliminary triage module, while the second decision threshold and the fuzzy decision interval width are retained by the optimization decision module.

[0081] The rejection mechanism action instructions generated by the optimization decision module include multiple control parameters for driving the rejection execution structure, and the control parameters include: the airflow injection angle of the rejection nozzle 104 dynamically determined according to the real-time posture data of the target tableware on the transmission line 101; the pulse air pressure intensity level set according to the degree of deviation between the fusion confidence score and the currently used second decision threshold, and at least three air pressure intensities are set, and the greater the degree of deviation, the higher the corresponding air pressure intensity level; the nozzle rejection trigger timing determined based on the production line transmission speed and the predicted position of the target tableware.

[0082] In an embodiment of the present invention, the optimal rejection mechanism action instructions are generated as follows: the real-time posture data of the target tableware on the transmission line 101 is obtained by a laser sensor installed above the transmission line 101, and the airflow injection angle is dynamically determined based on the posture data. For example, when the center of the tableware deviates from the center line of the conveyor belt by 5 mm, the injection angle is adjusted to 30 degrees with the running direction of the transmission line 101.

[0083] The intensity level of the pulsed air pressure is determined based on the degree of deviation between the fusion confidence score and the currently used second decision threshold. The deviation is set as the absolute value of the difference between the fusion confidence score and the second decision threshold. When the deviation is less than or equal to 0.1, the intensity level is level 1, and the air pressure is 0.2 MPa; when the deviation is less than or equal to 0.1, the intensity level is level 2, and the air pressure is 0.4 MPa; when the deviation is greater than 0.2, the intensity level is level 3, and the air pressure is 0.6 MPa. The timing of the rejection nozzle 104 is based on the production line speed (collected in real time by the encoder 107 of the transmission line 101) and the target tableware position. The rejection trigger time point is determined by the formula t = (Lx) / v, where L is the distance from the rejection nozzle 104 to the rejection zone, x is the distance from the current position of the target tableware to the rejection nozzle 104, v is the production line speed, and t is the trigger time.

[0084] The multi-level self-test program includes:

[0085] Perform signal integrity and operational stability diagnosis on the high-speed camera 103, the array of disturbance airflow nozzles 106, and the encoder 107 of the transmission line 101 that constitute the disturbance section, completing the equipment layer self-test;

[0086] During the equipment-level self-inspection or when no abnormality is found in the equipment-level self-inspection, drift detection analysis is performed on the initial confidence score sequence output by the prediction module, and stability analysis is performed on the fused confidence score sequence output by the disturbance response module to complete the algorithm-level self-inspection; when both the equipment-level self-inspection and the algorithm-level self-inspection confirm the existence of an abnormality and the cause of the abnormality is not located, a graded alarm signal is generated and pushed to the central monitoring platform to trigger manual intervention, and a pause operation instruction is sent to the production line control system.

[0087] The device layer self-test procedure includes:

[0088] The device layer self-test program specifically performs the following detection operations: sending a test signal to the high-speed camera 103 to analyze the temporal continuity and pixel integrity of the image frame; collecting the air pressure sensor readings of the disturbance airflow nozzle array 106 in real time and comparing them with the target air pressure intensity set by the optimization decision module; reading the displacement pulse of the encoder 107 of the transmission line 101, and synchronously correlating the tableware posture data captured by the disturbance response module to calculate the position offset; recording the triggering moment of the rejection instruction and the actual response time of the feedback signal of the rejection nozzle 104.

[0089] If no abnormalities are found during the device layer self-test, the algorithm layer self-test is performed, including:

[0090] Comparing and analyzing the real-time initial confidence score sequence currently output by the prediction module with a pre-established reference sequence of initial confidence scores during a stable operation benchmark period, and determining that the performance of the prediction model has drifted when the difference in statistical characteristics of the two sequences exceeds a preset drift threshold;

[0091] The real-time fusion confidence score sequence currently output by the disturbance response module is compared with the pre-established fusion confidence score reference sequence during the stable operation benchmark period for fluctuation analysis. When the fluctuation of the real-time sequence exceeds the preset fluctuation threshold or the corresponding indicator of the reference sequence, it is determined that the stability of the fusion confidence score output has decreased;

[0092] Performing pattern recognition on at least one of the real-time initial confidence score sequence output by the prediction module and the real-time fused confidence score sequence output by the disturbance response module, and determining that an abnormal pattern exists in the score output when detecting the presence of multiple consecutive identical score values ​​in the sequence, abnormal amplitude mutations in the score values, or frequent occurrence of extremely high or extremely low score values ​​within a short time window;

[0093] Pre-calibrated tableware benchmark samples are regularly input into the prediction module and the disturbance response module; the initial confidence score output by the prediction module and the fusion confidence score output by the disturbance response module are respectively compared with the preset benchmark scores corresponding to the benchmark samples; when the score deviation exceeds the preset allowable range, the abnormal information is recorded and an algorithm performance abnormality signal is sent to the optimization decision module.

[0094] The abnormal self-check module jointly analyzes the abnormal signals detected by the device layer, the real-time rejection rate fluctuation data and the score stability results output by the algorithm layer self-check; when at least two detection abnormalities in the device layer are met at the same time, the real-time rejection rate exceeds the dynamic safety threshold, and the algorithm layer self-check confirms that the initial confidence score of the prediction module drifts or the stability of the fusion confidence score decreases, a degradation instruction is sent to the optimization decision module and a protection mechanism is triggered: dynamically increase at least one of the first decision threshold and the second decision threshold, or suspend the execution of the current rejection action, or switch to a simplified decision strategy based on the historical average confidence score.

[0095] In an embodiment of the present invention, the device layer self-test program specifically performs the following test operations: sending a test signal to the high-speed camera 103, continuously sending 100 frames of test images, analyzing the temporal continuity of the image frames, requiring the standard deviation of the time interval between adjacent frames to be no more than 0.5ms, and pixel integrity by calculating the proportion of erroneous pixels in each frame of the image, requiring no more than 0.05%. Real-time acquisition of the air pressure sensor readings of the disturbance airflow nozzle array 106, 10 times per minute, comparing them with the target air pressure intensity set by the optimization decision module, calculating the average deviation, and allowing the average deviation to be no more than ±3%. Reading the displacement pulses of the encoder 107 of the transmission line 101, recording the number of pulses once per second, synchronously correlating the tableware posture data captured by the disturbance response module, calculating the average position offset per minute. In this embodiment, the allowable position offset is no more than ±0.5mm, and the average value of the rejection instruction response time is recorded to be less than or equal to 15ms to ensure the uniformity and operability of the device layer self-test standards.

[0096] Record the actual response time between the triggering moment of the rejection instruction and the feedback signal of the rejection nozzle 104, and calculate the average response time every hour, which is required to be no more than 15ms to ensure stable operation of the equipment.

[0097] The real-time initial confidence score sequence currently output by the prediction module is compared and analyzed with the initial confidence score reference sequence of the pre-established stable operation benchmark period (data from the normal production period of the past 7 days are selected), and the mean difference and variance difference between the two are calculated. When the mean difference exceeds the preset drift threshold (set to 0.1 in this embodiment) or the variance difference exceeds the preset drift threshold (set to 0.05 in this embodiment), it is determined that the prediction model performance has drifted.

[0098] The real-time fusion confidence score sequence currently output by the disturbance response module is compared and analyzed with a pre-established fusion confidence score reference sequence for a stable operation benchmark period, and the standard deviation of the real-time sequence is calculated. When the standard deviation of the real-time sequence exceeds the preset fluctuation threshold of the reference sequence standard deviation (set to 15% in this embodiment), it is determined that the stability of the fusion confidence score output has decreased.

[0099] Pattern recognition is performed on at least one of the real-time initial confidence score sequence output by the prediction module and the real-time fused confidence score sequence output by the disturbance response module. If five identical scores appear consecutively in the score sequence, or if the difference between adjacent scores exceeds a preset amplitude threshold (set to 0.2 in this embodiment), the score output is determined to have an abnormal pattern. A set of 100 pre-calibrated tableware benchmark samples is regularly (once a week) input into the prediction module and the disturbance response module. The initial confidence scores output by the prediction module and the fused confidence scores output by the disturbance response module are compared with the preset benchmark scores corresponding to the benchmark samples. If the score deviation exceeds ±0.1, the abnormality is recorded and a performance anomaly signal is sent to the optimization decision module.

[0100] The abnormality self-check module jointly analyzes the abnormal signals detected at the device layer, the real-time rejection rate fluctuation data, and the scoring stability results output by the algorithm layer self-check.

[0101] When at least two of the three tests, namely, abnormal continuity of image frame timing of the high-speed camera 103 in the equipment layer, deviation of the pressure sensor reading of the disturbance airflow nozzle 106 array exceeds the allowable value, and position offset of the encoder 107 of the transmission line 101 exceeds the standard, are continuously abnormal (lasting for more than 5 minutes), and the following conditions are met at the same time: the real-time rejection rate exceeds the dynamic safety threshold and the algorithm layer self-check confirms that the initial confidence score of the prediction module drifts (mean difference exceeds 0.15) or the stability of the fusion confidence score decreases (standard deviation exceeds 20% of the reference sequence), a degradation instruction is sent to the optimization decision module and the protection mechanism is triggered. In this embodiment, the dynamic safety threshold is calculated based on the average rejection rate in the past 24 hours, and is determined as the benchmark threshold for real-time monitoring by fluctuating the average value by ±20%: the first decision threshold and the second decision threshold are dynamically increased by a preset increment (0.1 each time in this embodiment), the current rejection action is suspended for 5 minutes, and at the same time, a simplified decision strategy based on the historical average confidence score (the average confidence score in the past hour) is switched to. After the simplified decision strategy runs for 10 minutes, if the abnormal situation has not been alleviated, a pause instruction is issued to the production line control system.

[0102] It should be noted that the above modules can be deployed in an industrial control platform with computing and control functions, which includes components such as a processor, a memory, a communication interface, and an execution control interface. Among them, the processor is used to execute algorithm programs such as image analysis, prediction model prediction, and dynamic parameter optimization; the memory is used to store image data, model parameters, historical rejection information, and control instructions; the communication interface establishes data communication with the image acquisition module (such as the high-speed camera 103), the execution module (such as the nozzle controller, which controls the rejection nozzle 104 and the disturbing airflow nozzle 106), the production line PLC system, etc.; the execution control interface is used to issue rejection instructions and collect rejection feedback signals. This is an existing technology and will not be described in detail.

[0103] In this embodiment, the flipping mechanism 102 is a robot arm that flips the bamboo tableware by means of a vacuum suction cup assembly provided on its surface. The working principles and usage procedures of the transmission line 101, flipping mechanism 102, high-speed camera 103, disturbed airflow nozzle 106, rejection nozzle 104, and encoder 107 are all prior art and will not be described in detail.

[0104] The optimization cycle of Bayesian optimization parameter adjustment is set to be executed every 30 minutes for example, and can be flexibly set according to needs in actual production process;

[0105] As attached Figure 2 As shown, the turning mechanism 102 is arranged on the transmission line 101, and high-speed cameras 103 are symmetrically arranged on both sides of the turning mechanism 102. The rectangular box 105 is arranged on the right side of the turning mechanism 102. The inside of the rectangular box 105 is installed with a disturbed airflow nozzle 106 and a high-speed camera 103. There are three groups of rejection nozzles 104, which are respectively arranged on the left side of the turning mechanism 102, between the turning mechanism 102 and the rectangular box 105, and on the right side of the rectangular box 105, so as to facilitate the transmission, detection and rejection of bamboo tableware.

[0106] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A big data bamboo tableware manufacturing system based on production process monitoring, including a transmission line and a turning mechanism, characterized in that: Also includes: An acquisition module, used to collect image data of tableware through high-speed cameras deployed on both sides of the turning mechanism; The prediction module is used to output the defect probability of each piece of tableware through a pre-trained prediction model based on the image data, and perform a weighted summation of the defect probabilities to generate an initial confidence score. The preliminary diversion module is used to set a first decision threshold based on the initial confidence score. If the initial confidence score is lower than the first decision threshold, the piece of tableware is judged as a high-probability defective product and is directly sent to the rejection zone. If the initial confidence score is equal to or higher than the threshold, the piece of tableware is transported to the disturbance section of the transmission line. A disturbance response module is configured to apply a preset airflow disturbance to tableware entering the disturbance section, obtain a dynamic image sequence of the tableware under the airflow disturbance from at least one perspective, calculate the tableware's dynamic stability index using optical flow analysis, and combine the tableware's initial image features with the dynamic stability index, process them using an ensemble learning algorithm, and output a fusion confidence score. An optimization decision module is used to set the second decision threshold and fuzzy decision interval parameters, determine the final rejection target based on the comparison result of the fusion confidence score and the fuzzy decision interval, obtain historical rejection data and real-time status data of the production line, and take minimizing the false rejection rate and ensuring that the defect detection rate meets the requirements as the optimization goal. Based on the historical rejection data and real-time status data of the production line, the first decision threshold, the second decision threshold and the fuzzy decision interval parameters are dynamically determined through the Bayesian optimization method, and the updated first decision threshold parameters are sent to the preliminary diversion module in real time. The second decision threshold and the fuzzy decision interval parameters are retained for self-use. Based on the fusion confidence score and the dynamically adjusted second decision threshold, the rejection mechanism action instructions are generated.

2. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 1 is characterized in that: Also includes: The abnormality self-check module is used to continuously monitor the tableware rejection rate within a set time window and set a dynamic safety threshold. The dynamic safety threshold is dynamically updated based on a historical rejection rate statistical model. When the real-time rejection rate continues to exceed the dynamic safety threshold for a predetermined period of time, a multi-level self-check program is automatically triggered.

3. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 1 is characterized in that: The pre-trained prediction model outputs an initial confidence score for each dish, including: The prediction model architecture adopts a deep residual network, whose main structure consists of four residual block groups, each consisting of two residual blocks. Two convolutional layers are set up in each residual block and a cross-layer identity mapping operation is performed to extract the surface defects and geometric deformation features of bamboo tableware. Horizontal connection channels are established at the output ends of the second and third residual block groups, respectively, and the bottom-level high-resolution feature maps and high-level strong semantic feature maps are input into the feature pyramid network. The resolution of the high-level feature maps is aligned through bilinear interpolation upsampling operations, and element-by-element addition and fusion are performed with the bottom-level feature maps of the corresponding levels to output the fused multi-scale feature maps. A hollow convolution layer is embedded in the third residual block group, and the feature weights of each region are dynamically adjusted by calculating the correlation weight matrix between the spatial positions of the feature maps. The output layer of the prediction model uses a multi-label classification framework: the fully connected layer contains 512 neurons and receives feature vectors weighted by self-attention. The output layer has the same number of independent sub-output layers as the number of defect types. Each sub-output layer uses a sigmoid activation function to output the probability value of the corresponding defect. The probability of each defect is weighted and summed according to the preset defect risk weight to generate an initial confidence score in the range of [0,1].

4. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 1 is characterized in that: The dynamic stability index of tableware is calculated by optical flow analysis, including: Based on a continuous image sequence of tableware under airflow disturbance obtained by the disturbance response module, the image acquisition frame rate is not less than 200 frames per second; the optical flow analysis method is used to track the displacement of at least five key geometric feature points preset on the surface of the tableware in the image sequence, wherein the key geometric feature points include the center point of the tableware, the point of maximum edge curvature, and the joint point, and the motion trajectory data of each feature point is generated; three characteristic parameters are extracted based on the motion trajectory data: the first is the maximum offset amplitude, which is obtained by calculating the straight-line distance of all feature points from the initial position to the maximum displacement point during the airflow disturbance, and taking the maximum value of the distance values ​​of all feature points as the maximum offset amplitude; the second is the posture recovery time, which is calculated from the moment the airflow disturbance ends, and the time interval in milliseconds during which the displacement of all feature points remains within the range of ±0.5 mm from the initial position for 100 milliseconds; the third is the resonance intensity coefficient, which is obtained by performing spectral analysis on the motion trajectory of the feature points and extracting the ratio of the energy of the main frequency component in the amplitude spectrum to the total energy to obtain the resonance intensity coefficient; and finally, the maximum offset amplitude, posture recovery time, and resonance intensity coefficient are combined to form a dynamic stability index.

5. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 1 is characterized in that: Dynamically determining decision parameters through Bayesian optimization methods includes: The objective function is defined as minimizing the false rejection rate, while setting a constraint that the defect detection rate is not lower than a predetermined standard value. Based on the fusion confidence score distribution in historical rejection data, the frequency of multiple defect types, and the real-time transmission speed characteristics of the production line, a Gaussian process regression algorithm is used to construct a mapping relationship between decision parameters and optimization objectives. The decision parameters include a first decision threshold, a second decision threshold, and the width of the fuzzy decision interval. Dynamic adjustment is achieved by iteratively executing the following process: periodically generating multiple sets of candidate parameter combinations within the range of decision parameter values, using the Gaussian process regression algorithm to predict the false rejection rate and defect detection rate corresponding to each combination, screening candidate combinations that meet the defect detection rate constraint and have a false rejection rate lower than the currently used parameters, and selecting the combination with the lower false rejection rate to update the parameter settings; outputting the updated first decision threshold to the preliminary diversion module, and outputting the second decision threshold and the fuzzy decision interval width to the optimization decision module.

6. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 1 is characterized in that: The rejection mechanism action instructions generated by the optimization decision module include multiple control parameters for driving the rejection execution structure, and the control parameters include: the rejection nozzle airflow injection angle dynamically determined according to the real-time posture data of the target tableware on the transmission line; the pulse air pressure intensity level set according to the degree of deviation between the fusion confidence score and the currently used second decision threshold, and at least three air pressure intensities are set, and the greater the degree of deviation, the higher the corresponding air pressure intensity level; the nozzle rejection trigger timing determined based on the production line transmission speed and the predicted position of the target tableware.

7. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 2 is characterized in that: The multi-level self-test program includes: Conduct signal integrity and operational stability diagnosis on the high-speed camera, disturbance airflow nozzle array, and transmission line encoder that constitute the disturbance section, completing equipment layer self-test; During the device-level self-test or when no abnormalities are found during the device-level self-test, drift detection analysis is performed on the initial confidence score sequence output by the prediction module, and stability analysis is performed on the fused confidence score sequence output by the disturbance response module to complete the algorithm-level self-test; When both the equipment-level self-inspection and the algorithm-level self-inspection confirm the existence of an abnormality and the cause of the abnormality has not been located, a graded alarm signal is generated and pushed to the central monitoring platform to trigger manual intervention, and a pause operation instruction is sent to the production line control system.

8. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 7 is characterized in that: If no abnormalities are found during the device layer self-test, the algorithm layer self-test is performed, including: Comparing and analyzing the real-time initial confidence score sequence currently output by the prediction module with a pre-established reference sequence of initial confidence scores during a stable operation benchmark period, and determining that the performance of the prediction model has drifted when the difference in statistical characteristics of the two sequences exceeds a preset drift threshold; The real-time fusion confidence score sequence currently output by the disturbance response module is compared with the pre-established fusion confidence score reference sequence during the stable operation benchmark period for fluctuation analysis. When the fluctuation of the real-time sequence exceeds the preset fluctuation threshold or the corresponding indicator of the reference sequence, it is determined that the stability of the fusion confidence score output has decreased; Performing pattern recognition on at least one of the real-time initial confidence score sequence output by the prediction module and the real-time fused confidence score sequence output by the disturbance response module, and determining that an abnormal pattern exists in the score output when detecting the presence of multiple consecutive identical score values ​​in the sequence, abnormal amplitude mutations in the score values, or frequent occurrence of extremely high or extremely low score values ​​within a short time window; Pre-calibrated tableware benchmark samples are regularly input into the prediction module and the disturbance response module; the initial confidence score output by the prediction module and the fusion confidence score output by the disturbance response module are respectively compared with the preset benchmark scores corresponding to the benchmark samples; when the score deviation exceeds the preset allowable range, the abnormal information is recorded and an algorithm performance abnormality signal is sent to the optimization decision module.

9. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 8 is characterized in that: The device layer self-test procedure includes: The device layer self-test program specifically performs the following detection operations: sending a test signal to the high-speed camera to analyze the timing continuity and pixel integrity of the image frame; collecting the air pressure sensor readings of the disturbance airflow nozzle array in real time and comparing them with the target air pressure intensity set by the optimization decision module; reading the transmission line encoder displacement pulse, synchronously correlating the tableware posture data captured by the disturbance response module to calculate the position offset; recording the triggering moment of the rejection instruction and the actual response time of the rejection nozzle feedback signal.

10. The big data bamboo tableware manufacturing system based on production process monitoring according to claim 7, characterized in that: The abnormal self-check module jointly analyzes the abnormal signals detected by the device layer, the real-time rejection rate fluctuation data and the score stability results output by the algorithm layer self-check; when at least two detection abnormalities in the device layer are met at the same time, the real-time rejection rate exceeds the dynamic safety threshold, and the algorithm layer self-check confirms that the initial confidence score of the prediction module drifts or the stability of the fusion confidence score decreases, a degradation instruction is sent to the optimization decision module and a protection mechanism is triggered: dynamically increase at least one of the first decision threshold and the second decision threshold, or suspend the execution of the current rejection action, or switch to a simplified decision strategy based on the historical average confidence score.

Images