Industrial vision-assisted warehouse automated storage and retrieval optimization method and system

Abstract

The present application relates to the technical fields of industrial automation, artificial intelligence and computer vision application, in particular to a warehouse automatic access optimization method and system assisted by industrial vision, comprising: collecting original perception data stream, obtaining feature vector of abnormal data area; obtaining physical reality probability and optical false probability through perception classification model processing; calculating visual confidence entropy; in response to the visual confidence entropy being greater than a preset entropy activation threshold, performing illusion evolution deduction to determine a physical confirmation index; executing hierarchical decision logic, including: in response to the visual confidence entropy being not greater than the preset entropy activation threshold, outputting a first access optimization action; in response to the visual confidence entropy being greater than the preset entropy activation threshold, outputting a second access optimization action; obtaining a physical reality label, and correcting the perception classification model and the model used in the illusion evolution deduction; the present application realizes active verification of passive perception failure, and improves decision reliability.

Description

Technical Field

[0001] This invention relates to the fields of industrial automation, artificial intelligence and computer vision application technology, specifically to an industrial vision-assisted automated storage and retrieval optimization method and system. Background Technology

[0002] In industrial vision-assisted light-out automated warehousing systems, vision is the only means of perception; the system needs to operate under a low signal-to-noise ratio and high-risk steady state caused by the continuous diffusion of optical interference particles. This extreme environment causes the perception system to frequently encounter optical noise clouds, or optical illusions, whose data characteristics are highly similar to real physical obstacles. Existing perception models struggle to reliably distinguish between the two, resulting in high uncertainty in perception results, i.e., high visual confidence entropy. This perceptual confusion exposes the system to catastrophic misalignment risks: misjudging an illusion as a real obstacle can lead to unnecessary avoidance or shutdown of access units, severely reducing operational efficiency; conversely, misjudging a real obstacle as an illusion could trigger a collision, causing a major safety accident. Therefore, ensuring that the system can accurately distinguish between optical illusions and real physical obstacles in this highly uncertain environment, and accordingly execute access optimization decisions that balance safety and efficiency, is a pressing technical problem that needs to be solved. Summary of the Invention

[0003] The purpose of this invention is to provide an industrial vision-assisted automated storage and retrieval optimization method and system for warehouses. This avoids the catastrophic misalignment risk caused by the perception system confusing optical illusions with real physical obstacles in a high-risk steady-state environment with low signal-to-noise ratio due to the continuous diffusion of optical interference particles. Furthermore, it ensures that the system accurately distinguishes between optical illusions and real physical obstacles in this high-uncertainty environment and accordingly executes storage and retrieval optimization decisions that balance safety and efficiency. Specifically, the technical solution of this invention is as follows: An industrial vision-assisted automated storage and retrieval optimization method for warehouses includes: Step 1: Collect raw sensing data stream and process the raw sensing data stream to obtain feature vectors of abnormal data regions; Step 2: Based on the feature vectors, the probability of physical authenticity and the probability of optical falsehood are obtained through a perceptual classification model. Step 3: Combine the probability of physical authenticity and the probability of optical falsehood to calculate the visual confidence entropy; Step 4: In response to the visual confidence entropy being greater than the preset entropy activation threshold, perform hallucination evolution deduction to determine the physical confirmation index; Step 5, execute the hierarchical decision-making logic, including: If the visual confidence entropy is not greater than the preset entropy activation threshold, then a normal path decision is performed based on the probability of physical authenticity and the probability of optical falsity, and the first access optimization action is output. In response to a visual confidence entropy greater than a preset entropy activation threshold, advanced path decision-making is performed based on the physical confirmation index and the preset physical confirmation threshold, and a second access optimization action is output. Step 6: In response to the second access optimization action being physical probing, obtain physical authenticity labels, and based on the physical authenticity labels and corresponding abnormal event data, correct the models used in the perception classification model and the hallucination evolution inference.

[0004] Preferably, processing the raw sensor data stream includes: Optical noise clouds can be detected in real time by performing inter-frame differencing on the raw sensing data stream or by using a pre-trained segmentation network to obtain feature vectors of anomalous data regions.

[0005] Preferably, a hallucination evolution deduction is performed to determine the physical confirmation index, including: The abnormal data region is input into a pre-trained hallucination model and a pre-trained real model to generate a predicted hallucination sequence and a predicted real sequence. Obtain real observation data; Based on real observation data and predicted hallucination sequences, calculate the hallucination simulation similarity score; Calculate the real simulation similarity score based on real observation data and predicted real sequences; By combining the similarity scores of hallucination simulations and real simulations, and processing them using a normalized ratio formula, a physical confirmation index is determined.

[0006] Preferably, calculating the similarity score between the hallucination simulation and the real simulation includes: The calculation is performed using a structural similarity algorithm or a normalized cross-correlation score algorithm.

[0007] Preferred, conventional path decisions include: In response to the fact that the probability of physical authenticity is greater than the probability of optical falsehood, the first access optimization action is to avoid standard obstacles. In response to the fact that the probability of physical authenticity is not greater than the probability of optical falsity, the first access optimization action is to continue executing the data-driven optimal path solution.

[0008] Preferred, advanced path decision-making includes: In response to the physical confirmation index being within a preset ambiguity range, the second access optimization action is changed to perform a perception holding action.

[0009] Preferred advanced path decision-making also includes: In response to a physical confirmation index exceeding a preset physical confirmation threshold, or a timeout of the perception hold action, the second access optimization action is to perform a physical probe.

[0010] Preferred, modified models include: Pack the physical authenticity labels and anomalous event data into training samples; Gradient descent is used to fine-tune the perceptual classification model, hallucination model, and real model online using training samples.

[0011] An industrial vision-assisted automated storage and retrieval optimization system for warehouses includes: The data processing module is used to collect and process the raw sensing data stream and obtain the feature vectors of abnormal data regions. The perceptual quantization module is used to obtain the probability of physical authenticity and the probability of optical falsehood based on feature vectors through a perceptual classification model. The entropy calculation module is used to calculate the visual confidence entropy by combining the probability of physical authenticity and the probability of optical falsehood. The evolutionary deduction module is used to perform hallucination evolution deduction and determine the physical confirmation index in response to the visual confidence entropy being greater than the preset entropy activation threshold. The decision execution module is used to execute hierarchical decision logic, including: The regular path unit is used to respond to the fact that the visual confidence entropy is not greater than the preset entropy activation threshold, and then to perform a regular path decision based on the probability of physical authenticity and the probability of optical falsehood, and output the first access optimization action. The advanced path unit is used to respond to a visual confidence entropy greater than a preset entropy activation threshold by performing advanced path decisions based on the physical confirmation index and the preset physical confirmation threshold, and outputting a second access optimization action. The model correction module is used to respond to the second access optimization action as a physical probe, obtain physical authenticity labels, and correct the models used in the perception classification model and evolution inference module based on the physical authenticity labels and the corresponding abnormal event data.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention quantifies the degree of confusion between real obstacles and optical illusions by calculating the probability of physical authenticity and the probability of optical illusion, and combines the two to calculate the visual confidence entropy, thus providing a reliable basis for subsequent hierarchical decision-making. 2. This invention employs a hierarchical decision-making logic; when the perceived confidence is high, it executes conventional path decision-making to ensure high access efficiency; when the perceived confidence is low, i.e., highly ambiguous, it switches to advanced path decision-making to prioritize safety, thus achieving a dynamic balance between efficiency and safety. 3. This invention innovatively performs hallucination evolution deduction for highly uncertain events; by comparing the predicted hallucination sequence with the real sequence and actual observation data, a physical confirmation index is determined, realizing active verification of passive perception failure and improving decision reliability; 4. This invention has closed-loop learning capability; after performing physical exploration, it obtains physical authenticity labels, and uses these labels and abnormal event data to correct the perception classification model and evolutionary inference model online, so that the system can continuously adapt to unknown interference and improve long-term robustness. Attached Figure Description

[0013] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0015] Example 1: Please see Figure 1 An industrial vision-assisted automated storage and retrieval optimization method for warehouses includes: Step 1: Collect raw sensing data stream and process the raw sensing data stream to obtain feature vectors of abnormal data regions; Step 2: Based on the feature vectors, the probability of physical authenticity and the probability of optical falsehood are obtained through a perceptual classification model. Step 3: Combine the probability of physical authenticity and the probability of optical falsehood to calculate the visual confidence entropy; Step 4: In response to the visual confidence entropy being greater than the preset entropy activation threshold, perform hallucination evolution deduction to determine the physical confirmation index; Step 5, execute the hierarchical decision-making logic, including: If the visual confidence entropy is not greater than the preset entropy activation threshold, then a normal path decision is performed based on the probability of physical authenticity and the probability of optical falsity, and the first access optimization action is output. In response to a visual confidence entropy greater than a preset entropy activation threshold, advanced path decision-making is performed based on the physical confirmation index and the preset physical confirmation threshold, and a second access optimization action is output. Step 6: In response to the second access optimization action being physical probing, obtain physical authenticity labels, and based on the physical authenticity labels and corresponding abnormal event data, correct the models used in the perception classification model and the hallucination evolution inference.

[0016] This embodiment provides an industrial vision-assisted automated storage and retrieval optimization method; the method is applied to a light-out automated storage system, an environment characterized by the vision system being the only means of perception, and the system operating under a high-risk steady state with low signal-to-noise ratio caused by the continuous diffusion of optical interference particles; Step 1: Collect raw sensing data stream and process the raw sensing data stream to obtain feature vectors of abnormal data regions; The purpose of collecting raw sensing data streams is to obtain raw information about the environment. In this embodiment, sensors such as depth cameras or LiDAR are deployed in the warehouse environment or on AGVs to collect real-time raw sensing data streams, denoted as... ; The purpose of processing the raw sensing data stream is to segment potential abnormal events from the continuous data stream; this embodiment achieves this through sudden abnormal region segmentation; specifically, the system segments the data stream... By performing inter-frame differencing or utilizing a pre-trained segmentation network, optical noise clouds that the system has not learned from, caused by factors such as impulse leakage, can be detected in real time; the output of this processing step is the anomalous data region. and its corresponding eigenvectors ; Step 2: Based on the feature vectors, the probability of physical authenticity and the probability of optical falsehood are obtained through a perceptual classification model. The purpose of processing feature vectors through a perceptual classification model is to segment out anomaly regions. A preliminary property assessment is performed; in this embodiment, the feature vector obtained in step 1... It is fed into a pre-trained perceptual classification model; A perceptual classification model refers to a classifier trained on historical data, which distinguishes between real solid obstacles and optical illusions; it is obtained by supervised training using historical datasets containing simulated dust, fog, and real obstacles. Obtaining the probability of physical authenticity and the probability of optical falsehood: This model processes... Then, output two core probability values: The probability of physical authenticity is denoted as : i.e., data The probability of being judged as physically real (Real); Optical falsehood probability is denoted as : i.e., data The probability of being identified as an optical falsehood (Illusion); in, ; Step 3: Combine the probability of physical authenticity and the probability of optical falsehood to calculate the visual confidence entropy; The purpose of calculating visual confidence entropy is to provide a quantitative indicator for the uncertainty of a perceptual system, that is, to quantify the degree to which the system confuses illusion with reality at the current level; in this embodiment, visual confidence entropy is denoted as... It is calculated using the following formula derived from Shannon's information entropy:

[0017] in Visual confidence entropy is a dimensionless scalar, the unit of which can be regarded as bits. It is calculated by this formula and its function is to measure the degree of confusion of the system with the current perception. The probability of physical authenticity is a dimensionless probability value, obtained from the output of the perceptual classification model in step 2. Let be the probability of optical falsehood, which is a dimensionless probability value, and This is obtained from the output of the perceptual classification model in step 2; The technical motivation behind this formula is that when and When all values ​​are close to 0.5, the system is at its most confused. Reaching the maximum value of 1 indicates the highest risk of catastrophic misalignment; The calculation results will serve as a key input for subsequent hierarchical decision-making; Step 4: In response to the visual confidence entropy being greater than the preset entropy activation threshold, perform hallucination evolution deduction to determine the physical confirmation index; The system initiates active verification when the visual confidence entropy exceeds a preset entropy activation threshold. ; Entropy activation threshold This refers to a scalar used to determine whether the degree of perceptual ambiguity reaches a dangerous level; its source is based on offline simulation and risk assessment presets, such as... ;when When this occurs, it indicates that the perception is highly ambiguous, and the system considers step 2 to be... The results are unreliable; Performing hallucination evolution deduction to determine the physical confirmation index: its purpose is to actively test the physical authenticity of the data when passive perception step 2 fails; this process is detailed in Example 3 by calling two parallel generative models, one simulating hallucination evolution and the other simulating the evolution of real objects, to predict abnormal regions. In the future Changes over time, and with The data was compared with actual observation data after a certain time; finally, a physical confirmation index was calculated using a normalized ratio formula, denoted as [index name missing]. The index A value range close to [0, 1] is more consistent with real object models when used to quantify observation data. Or an optical illusion model ; Step 5: Execute the hierarchical decision-making logic; The purpose of this step is to determine the degree of uncertainty in perception. and the results of active verification To achieve the optimal balance between efficiency and safety; If the visual confidence entropy is not greater than the preset entropy activation threshold, then a conventional path decision is performed based on the probability of physical authenticity and the probability of optical falsehood, and the first access optimization action is output: This is Level 1: Normal path; the trigger condition is... This indicates clear perception and high confidence; the system trusts step 2. and The judgment; the conventional path decision performed at this time is detailed in Example 5, which will output the first access optimization action, for example: if If determined to be a genuine obstacle, standard avoidance procedures are followed; if If the anomaly is determined to be a hallucination, the AGV will continue to execute the optimal path and ignore the anomaly. If the visual confidence entropy exceeds a preset entropy activation threshold, then advanced path decision-making is performed based on the physical confirmation index and the preset physical confirmation threshold, outputting a second access optimization action: This is either Level 2: Active Verification or Level 3: Physical Probe; the trigger condition is... This indicates that the perception is highly ambiguous; at this point, the system no longer trusts it. Instead, it relies on the physical confirmation index calculated in step 4. And a preset physical confirmation threshold. ; Physical confirmation threshold It refers to a tool used for judgment Whether the value has a significant bias; its source is based on the accuracy preset of offline simulation and generative models, for example... ; The advanced path decision executed at this point will output a second access optimization action; for example: if If the value falls within a fuzzy range, such as between 0.3 and 0.7, then the system will either remain stationary or crawl, waiting for the target to complete its journey. Convergence; if If a value conflict occurs or verification times out, a physical probe is performed. Step 6: In response to the second access optimization action being physical probing, obtain physical authenticity labels, and based on the physical authenticity labels and corresponding abnormal event data, correct the models used in the perception classification model and the hallucination evolution inference. In response to the second access optimization action, physical probing occurs when the system decides to execute level three: physical probing. For example, the AGV actively slows down and uses the actuator to lightly touch the abnormal area. This step is activated at the boundary; Obtaining Physical Authenticity Labels: By analyzing the interaction results of physical probing, the system obtains information about abnormal areas. The ultimate physical truth value; for example: if the AGV actuator image passes through ,but Confirmed as an optical illusion, physical reality label Set to 0; if If the image is pushed or the actuator stops at the boundary, then Confirmed as physically real, tagged Set to 1; Based on physical authenticity labels and corresponding anomalous event data, the perceptual classification model and the model used in hallucination evolution deduction are revised: the aim is to achieve closed-loop learning and self-evolution of the perceptual model; the system will display the complete data stream of this anomalous event. Include With the truth label just obtained Packaged to form a new high-value training sample The system uses this sample to analyze the perceptual classification model in step 2 and the generative model used in the hallucination evolution deduction in step 4, specifically the hallucination model. and real model Online fine-tuning or incremental training is described in detail in Example 8; The method described in this embodiment constructs a complete technical closed loop, from quantifying uncertainty to active verification, and then to resilient decision-making and closed-loop correction. It enables the warehousing system to effectively distinguish between optical illusions and real obstacles in high-risk, low-signal-to-noise-ratio dark environments, avoiding catastrophic misalignment caused by perceptual confusion. Through hierarchical decision-making, the system can effectively distinguish between optical illusions and real obstacles when perception is clear. Low hold and high access efficiency, even when perception is ambiguous. The system actively switches to a safety-first verification strategy; more importantly, by obtaining truth values ​​and correcting the model through physical probing, the system can learn from events with the highest uncertainty, continuously adapt to unknown disturbances that constantly arise in the environment, and achieve long-term perception robustness and system resilience.

[0018] Example 2: Processing the raw sensor data stream includes: Optical noise clouds can be detected in real time by performing inter-frame differencing on the raw sensing data stream or by using a pre-trained segmentation network to obtain feature vectors of anomalous data regions.

[0019] This embodiment is a concretization of step 1 in embodiment 1, which processes the raw sensing data stream. In this embodiment, processing the raw sensing data stream specifically refers to the stage of segmenting sudden anomaly regions; in order to process the raw data stream... Real-time detection and segmentation of abnormal regions This embodiment employs at least one method: by performing inter-frame differencing on the original sensing data stream, this method is suitable for detecting dynamic optical noise clouds; by calculating the temporal differences between consecutive data frames, whether images or point clouds, regions whose shape or position changes rapidly due to reasons such as pulse leakage can be quickly located, i.e. Alternatively, a pre-trained segmentation network can be used. This approach is suitable for detecting optical noise clouds with relatively stable morphology or insignificant inter-frame differences. The system employs a pre-trained segmentation network, such as U-Net or its variants, specifically trained for various types of optical noise in dark environments, such as dust and fog, and can directly analyze single-frame data. Mid-segment ; The abnormal data region was located and segmented using the methods described above. Then, the system extracts the feature vector of the region. For example, it could be a depth distribution histogram of the region, point cloud geometric features, or activation values ​​of intermediate layers of a segmentation network, for use in step 2 of Example 1; This embodiment ensures real-time and efficient detection of optical noise clouds by providing two specific techniques: inter-frame difference or pre-trained segmentation networks. These two methods are optimized for dynamic and static interference, respectively, guaranteeing the detection of abnormal data areas. It can be captured quickly and accurately, providing timely and high-quality input for the subsequent perception and classification step 2, thereby improving the sensitivity and speed of the entire optimization method in responding to abnormal events.

[0020] Example 3: Perform a hallucination evolution deduction to determine the physical confirmation index, including: The abnormal data region is input into a pre-trained hallucination model and a pre-trained real model to generate a predicted hallucination sequence and a predicted real sequence. Obtain real observation data; Based on real observation data and predicted hallucination sequences, calculate the hallucination simulation similarity score; Calculate the real simulation similarity score based on real observation data and predicted real sequences; By combining the similarity scores of hallucination simulations and real simulations, and processing them using a normalized ratio formula, a physical confirmation index is determined.

[0021] This embodiment is a concretization of step 4 in embodiment 1, which involves performing hallucination evolution deduction to determine the physical confirmation index; when The process is initiated at that time: The abnormal data region is input into a pre-trained hallucination model and a pre-trained real model to generate predicted hallucination sequences and predicted real sequences; the system calls two parallel, pre-trained generative models; the pre-trained hallucination model... Its function is to simulate As an optical noise cloud Evolution over time; pre-trained real model Its function is to simulate As a real solid obstacle Evolution over time; both models are based on physics engine simulation data in a dark environment, such as dust diffusion in fluid dynamics simulations and obstacles in rigid body dynamics simulations, for pre-training; and the current anomalous data areas. As input at time t, respectively, they are fed into and ; Output the predicted hallucination sequence ; Output the predicted true sequence ; System waiting Time, Acquisition The actual observation data at time is denoted as ; System Calculation and The similarity between them is used to obtain the hallucination simulation similarity score. ; And calculate and The similarity between them is used to obtain a true simulation similarity score. ;about The specific calculation of the function is described in detail in Example 4; For comparison and The system calculates the physical confirmation index using a normalized ratio formula to assess the strength of evidence. :

[0022] in The physical confirmation index is a dimensionless scalar, calculated using this formula. To realistically simulate the similarity score, it is a dimensionless scalar with a value range of [0, 1], and is calculated by the previous steps; The similarity score for the hallucination simulation is a dimensionless scalar with a value range of [0, 1], which is calculated from the previous steps. To prevent zero positive numbers, a preset, extremely small positive number is used, such as... Its origin is a standard setting in numerical calculations, used to prevent the denominator from being zero; Should A value close to 1 indicates that the observed evolutionary trend is similar to the real object model. Highly consistent ; A value close to 0 indicates the data evolution trend and the optical illusion model. Highly consistent ; This embodiment details the process of actively testing the physical authenticity of data; by simulating both hallucination and reality in parallel and comparing the results with future real observations, this method transcends the limitations of passive perception; physical confirmation index. The calculation is for the system in the initial perception of high ambiguity. In high-level states, it provides crucial, evolutionary trend-based evidence for decision-making, greatly improving the reliability of decisions under high uncertainty.

[0023] Example 4: Calculate the similarity score for the hallucination simulation and the similarity score for the real simulation, including: The calculation is performed using a structural similarity algorithm or a normalized cross-correlation score algorithm.

[0024] This embodiment is a concretization of the similarity score calculation steps in Embodiment 3; Calculating similarity scores for hallucination simulations And calculate the similarity score between the real simulation and the real simulation. At that time, the method used The function must be a similarity metric that is as high as possible, and its results should be normalized to [value missing]. Closed intervals; the use of error metrics such as L2 distance and Euclidean distance, which are considered better the lower they are, is prohibited; In this embodiment, The function specifically employs either a structural similarity algorithm or a normalized cross-correlation scoring algorithm. The structural similarity algorithm SSIM is used when both the observed data and the predicted sequence are images. This algorithm compares the brightness, contrast, and structure of two images to obtain a similarity score between -1 and 1, typically within the range [0, 1], which is considered better the higher the score. The normalized cross-correlation scoring algorithm NCC can be used when the data is point clouds or images. This algorithm calculates the true observation... With predicted sequence Right now or The correlation between them is normalized to obtain a score in the range of [0, 1] or [-1, 1], which also meets the requirement that the higher the score, the better. This embodiment ensures a high similarity score by limiting the use of normalized similarity metrics such as SSIM or NCC, where higher scores are preferred. and The physical meaning and mathematical validity of SSIM; compared to noise-sensitive error metrics such as L2 distance, SSIM or NCC are better able to capture the structural or trend consistency between observed data and predicted sequences, thereby improving... The accuracy and robustness of index calculations make subsequent high-level path decisions more reliable.

[0025] Example 5: Conventional path decision-making includes: In response to the fact that the probability of physical authenticity is greater than the probability of optical falsehood, the first access optimization action is to avoid standard obstacles. In response to the fact that the probability of physical authenticity is not greater than the probability of optical falsity, the first access optimization action is to continue executing the data-driven optimal path solution.

[0026] This embodiment is a concretization of the conventional path decision-making step 5 in Embodiment 1; Conventional path decision-making Triggered under conditions of clear perception, the system at this point trusts the output of step 2. and The specific decision for the first access optimization action is as follows: The response is that the probability of physical authenticity is greater than the probability of optical falsehood, i.e., when At that time, the system identifies abnormal areas. For physical realism; the AGV performs standard obstacle avoidance actions, such as stopping, waiting for obstacles to be removed, or replanning a detour route; In response to the probability of physical authenticity not being greater than the probability of optical falsehood, i.e. when At that time, the system identifies abnormal areas. This is an optical illusion; the first optimization action of the AGV is to ignore this illusion and continue to execute the data-driven optimal path solution, maintaining the original efficient access operation. This embodiment clarifies the system's operational logic in a state of perceived confidence; by... and The comparison results are directly mapped to two drastically different actions: avoidance or ignoring. This method achieves low-risk... It maximizes access efficiency in low-level conditions; it ensures that the AGV will not make unnecessary stops or detours due to optical illusions, while guaranteeing a reliable response to real obstacles.

[0027] Example 6: Advanced path decision-making, including: In response to the physical confirmation index being within a preset ambiguity range, the second access optimization action is changed to perform a perception holding action.

[0028] This embodiment describes a specific scenario, level two, of the advanced path decision-making step 5 in embodiment 1; Advanced path decision-making in It is triggered under conditions of high perceptual ambiguity; at this time, the system depends on... Make decisions based on value; In response to the physical confirmation index falling within a preset ambiguity range, the second access optimization action is changed to a perception hold action; the preset ambiguity range is defined based on the preset physical confirmation threshold in Example 1. As determined by 0.7; this interval is defined as ,For example The trigger condition is when and Sometimes, or The calculation is not yet complete; this state represents the worst-case scenario, i.e., passive perception. Highly ambiguous, while actively evolving and deducing Unable to clearly distinguish between reality and illusion The value is around 0.5, with no obvious tendency; the second access optimization action: the system immediately abandons the optimal path solution and instead executes the perception-maintaining action; this action includes: the AGV stopping in place, or crawling at the minimum safe speed; at the same time, the system continues to run the illusion evolution deduction process, that is, continuously calculates and updates. Value, until Converging to outside the fuzzy interval or ; This embodiment defines the core safety strategy of the system under extreme uncertainty. When both passive perception and active inference fail, the perception-maintaining action forces the system to prioritize safety above all else, sacrificing efficiency. This strategy of waiting in place and continuously verifying avoids catastrophic errors by the AGV under highly confusing data, such as colliding with real obstacles or permanently stopping due to hallucinations. It buys the system valuable time to obtain clearer decision evidence, i.e., convergence. value.

[0029] Example 7: Advanced path decision-making also includes: In response to a physical confirmation index exceeding a preset physical confirmation threshold, or a timeout of the perception hold action, the second access optimization action is to perform a physical probe.

[0030] This embodiment describes another specific scenario of advanced path decision-making, level three, which can be followed by a perception-holding action or triggered under other conflict conditions. Advanced path decision-making in Execute under the following conditions; In response to a physical confirmation index exceeding a preset physical confirmation threshold, or a timeout in the perception hold action, the second access optimization action is changed to performing a physical probe; this is Level 3: Physical Probe; the triggering conditions include: the physical confirmation index exceeding a preset physical confirmation threshold. This is a situation of perception-deduction conflict; that is, passive perception. Very high, for example and Both are close to 0.5, or It's also very high, but actively extrapolating... However, it strongly leans towards real objects; or the perception holding action times out, which is a verification failure situation; that is, the system executes the perception holding action of Example 6, but within the preset timeout period, such as 5 seconds, The value still failed to converge to outside the fuzzy interval; In both cases, perception and simulation are highly conflicting or have been proven ineffective; the system must adopt a suboptimal solution based on physical trial and error, which is more likely to confirm the authenticity of the data. The second access optimization action is to perform a physical probe; this action is a confirmatory physical interaction, such as the AGV actively slowing down and gently touching an abnormal area. The system will obtain the final truth label by monitoring whether the interactive image crosses or pushes / stops. ; This embodiment defines the system's final decision-making mechanism; when all data-based sensing and simulation methods are passively classified... and active simulation When high levels of ambiguity or conflict cannot be resolved, this method proactively upgrades the decision-making process, exchanging controllable physical interaction for 100% accurate truth values. This physical probing action is not only the ultimate safe means to resolve the current high-risk dilemma, but also the only way to trigger subsequent closed-loop correction of the model, and a key step for the system to achieve long-term learning and adaptability.

[0031] Example 8: The model has been revised, including: Pack the physical authenticity labels and anomalous event data into training samples; Gradient descent is used to fine-tune the perceptual classification model, hallucination model, and real model online using training samples.

[0032] This embodiment is a concretization of the modified model in step 6 of embodiment 1; The process of revising the model is completed after the physics trial is executed and the truth value is obtained. Start after; The physical authenticity labels and anomalous event data are packaged into training samples; the system obtains the physical authenticity labels. Right now For the sake of truth, Complete abnormal event data corresponding to the hallucination and this event. The two were packaged into a new high-value training sample: ; Gradient descent is used to fine-tune the perceptual classification model, hallucination model, and real model online using training samples; the system then uses this new sample. The following models were fine-tuned or incrementally trained online: the perceptual classification model from Example 1, step 2, and the hallucination model. From Example 3 and the real model From Example 3, the model weights are adjusted using gradient descent or its variants, such as Adam and SGD.

[0033] in For the updated model weights, and The same dimensions are the result of this update; The original weights of the model refer to the current weight parameters of model M, i.e., the perception model or the generative model. The model learning rate is a preset scalar hyperparameter, which is determined by tuning during offline model training. The gradient of the model's loss function is the loss function. Weights The gradient; This refers to the model loss function, such as the cross-entropy loss function. The model refers to the modified perceptual or generative model. This is abnormal event data, which is the input of the packaged training samples; The true value label is the label of the packaged training sample; The technical motivation behind this formula is to achieve closed-loop learning of the perceptual model by adjusting the model weights using newly acquired ground truth samples in order to reduce future losses, i.e., errors, on similar events. This embodiment employs gradient descent to fine-tune the relevant model online, transforming the true values ​​obtained from physical trials into the model's experience. This closed-loop correction mechanism ensures that the system can learn and evolve from the highest uncertainty events it encounters, continuously improving its ability to identify unknown interference in dark environments, thus enhancing the accuracy and predictive power of the perceptual classification model. The accuracy of the model is the core guarantee for achieving long-term perception robustness and adaptive capability of the system.

[0034] Example 9: Please see Figure 2 An industrial vision-assisted automated storage and retrieval optimization system for warehouses, comprising: The data processing module is used to collect and process the raw sensing data stream and obtain the feature vectors of abnormal data regions. The perceptual quantization module is used to obtain the probability of physical authenticity and the probability of optical falsehood based on feature vectors through a perceptual classification model. The entropy calculation module is used to calculate the visual confidence entropy by combining the probability of physical authenticity and the probability of optical falsehood. The evolutionary deduction module is used to perform hallucination evolution deduction and determine the physical confirmation index in response to the visual confidence entropy being greater than the preset entropy activation threshold. The decision execution module is used to execute hierarchical decision logic, including: The regular path unit is used to respond to the fact that the visual confidence entropy is not greater than the preset entropy activation threshold, and then to perform a regular path decision based on the probability of physical authenticity and the probability of optical falsehood, and output the first access optimization action. The advanced path unit is used to respond to a visual confidence entropy greater than a preset entropy activation threshold by performing advanced path decisions based on the physical confirmation index and the preset physical confirmation threshold, and outputting a second access optimization action. The model correction module is used to respond to the second access optimization action as a physical probe, obtain physical authenticity labels, and correct the models used in the perception classification model and evolution inference module based on the physical authenticity labels and the corresponding abnormal event data.

[0035] This embodiment provides an industrial vision-assisted automated storage and retrieval optimization system for warehouses. This system can be implemented as a combination of hardware, software, or firmware, including: The data processing module is used to connect to sensors such as depth cameras to acquire raw sensor data streams. It processes the raw sensor data stream using built-in inter-frame difference or segmentation networks to obtain feature vectors of abnormal data regions. ; The perceptual quantization module, which deploys a perceptual classification model, is used to receive... and output the probability of physical authenticity. and the probability of optical falsehood ; Entropy calculation module, this module is used to receive and And according to Shannon's entropy formula Calculate the visual confidence entropy ; Evolutionary deduction module, which deploys the hallucination model and real model It is used in response to visual confidence entropy. Entropy calculation module values ​​are greater than the preset entropy activation threshold When activated, it performs hallucination evolution deduction, including generating sequences. , to obtain observations And use algorithms such as SSIM to calculate Finally, the physical confirmation index is determined and output. ; The decision execution module, which is the system's decision-making center, is used to execute hierarchical decision-making logic; it receives... and As input, and includes: a regular path unit, used in At that time, based on Perform a standard path decision and output the first access optimization action to the AGV controller; the advanced path unit is used to... At that time, based on and Perform advanced path decisions and output secondary access optimization actions, such as sense hold or physical probe, to the AGV controller; The model correction module, which responds to the second access optimization action as a physics probe issued by the decision execution module, obtains the physics accuracy label from the decision execution module. And obtain the corresponding abnormal event data from the data processing module. According to the gradient descent method, the weights are corrected and updated. The perceptual classification model deployed in the perceptual quantization module and the model deployed in the evolutionary inference module ; The system in this embodiment fully implements the optimization method proposed in this invention through the modular functional division and collaborative work described above. The system architecture is clear, forming a robust and adaptive automatic access optimization system from the data input and data processing module to uncertainty quantification, entropy calculation, active verification evolution and deduction, resilient decision execution and closed-loop learning model correction. The system can effectively cope with optical interference in dark environments, balance access efficiency and operational safety, and has the ability to continuously evolve.

[0036] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An industrial vision-assisted automated storage and retrieval optimization method for warehouses, characterized in that, include: Step 1: Collect raw sensing data stream and process the raw sensing data stream to obtain feature vectors of abnormal data regions; Step 2: Based on the feature vectors, the probability of physical authenticity and the probability of optical falsehood are obtained through a perceptual classification model. Step 3: Combine the probability of physical authenticity and the probability of optical falsehood to calculate the visual confidence entropy; Step 4: In response to the visual confidence entropy being greater than the preset entropy activation threshold, perform hallucination evolution deduction to determine the physical confirmation index; Step 5, execute the hierarchical decision-making logic, including: If the visual confidence entropy is not greater than the preset entropy activation threshold, then a normal path decision is performed based on the probability of physical authenticity and the probability of optical falsity, and the first access optimization action is output. In response to a visual confidence entropy greater than a preset entropy activation threshold, advanced path decision-making is performed based on the physical confirmation index and the preset physical confirmation threshold, and a second access optimization action is output. Step 6: In response to the second access optimization action being physical probing, obtain physical authenticity labels, and based on the physical authenticity labels and corresponding abnormal event data, correct the models used in the perception classification model and the hallucination evolution inference.

2. The industrial vision-assisted automated storage and retrieval optimization method for warehouses according to claim 1, characterized in that, Processing the raw sensor data stream includes: Optical noise clouds can be detected in real time by performing inter-frame differencing on the raw sensing data stream or by using a pre-trained segmentation network to obtain feature vectors of anomalous data regions.

3. The industrial vision-assisted automated storage and retrieval optimization method for warehouses according to claim 1, characterized in that, Perform a hallucination evolution deduction to determine the physical confirmation index, including: The abnormal data region is input into a pre-trained hallucination model and a pre-trained real model to generate a predicted hallucination sequence and a predicted real sequence. Obtain real observation data; Based on real observation data and predicted hallucination sequences, calculate the hallucination simulation similarity score; Calculate the real simulation similarity score based on real observation data and predicted real sequences; By combining the similarity scores of hallucination simulations and real simulations, and processing them using a normalized ratio formula, a physical confirmation index is determined.

4. The industrial vision-assisted automated storage and retrieval optimization method for warehouses according to claim 3, characterized in that, Calculate the similarity score for the hallucination simulation and the similarity score for the real simulation, including: The calculation is performed using a structural similarity algorithm or a normalized cross-correlation score algorithm.

5. The industrial vision-assisted automated storage and retrieval optimization method for warehouses according to claim 1, characterized in that, Conventional path decision-making includes: In response to the fact that the probability of physical authenticity is greater than the probability of optical falsehood, the first access optimization action is to avoid standard obstacles. In response to the fact that the probability of physical authenticity is not greater than the probability of optical falsity, the first access optimization action is to continue executing the data-driven optimal path solution.

6. The industrial vision-assisted automated storage and retrieval optimization method for warehouses according to claim 1, characterized in that, Advanced path decision-making, including: In response to the physical confirmation index being within a preset ambiguity range, the second access optimization action is changed to perform a perception holding action.

7. The industrial vision-assisted automated storage and retrieval optimization method for warehouses according to claim 1 or 6, characterized in that, Advanced path decision-making also includes: In response to a physical confirmation index exceeding a preset physical confirmation threshold, or a timeout of the perception hold action, the second access optimization action is to perform a physical probe.

8. The industrial vision-assisted automated storage and retrieval optimization method for warehouses according to claim 1, characterized in that, The model has been revised, including: Pack the physical authenticity labels and anomalous event data into training samples; Gradient descent is used to fine-tune the perceptual classification model, hallucination model, and real model online using training samples.

9. An industrial vision-assisted automated storage and retrieval optimization system for warehouses, applied to the industrial vision-assisted automated storage and retrieval optimization method for warehouses as described in any one of claims 1-8, characterized in that, include: The data processing module is used to collect and process the raw sensing data stream and obtain the feature vectors of abnormal data regions. The perceptual quantization module is used to obtain the probability of physical authenticity and the probability of optical falsehood based on feature vectors through a perceptual classification model. The entropy calculation module is used to calculate the visual confidence entropy by combining the probability of physical authenticity and the probability of optical falsehood. The evolutionary deduction module is used to perform hallucination evolution deduction and determine the physical confirmation index in response to the visual confidence entropy being greater than the preset entropy activation threshold. The decision execution module is used to execute hierarchical decision logic, including: The regular path unit is used to respond to the fact that the visual confidence entropy is not greater than the preset entropy activation threshold, and then to perform a regular path decision based on the probability of physical authenticity and the probability of optical falsehood, and output the first access optimization action. The advanced path unit is used to respond to a visual confidence entropy greater than a preset entropy activation threshold by performing advanced path decisions based on the physical confirmation index and the preset physical confirmation threshold, and outputting a second access optimization action. The model correction module is used to respond to the second access optimization action as a physical probe, obtain physical authenticity labels, and correct the models used in the perception classification model and evolution inference module based on the physical authenticity labels and the corresponding abnormal event data.

Images