A motor sound quality self-adaptive detection method, system, device and medium based on uncertainty value

Abstract

The application discloses a motor sound quality adaptive detection method and system based on uncertainty values, equipment and a medium, mainly related to the technical field of motor sound quality adaptive detection, to solve the problems of long model training period, high cost in the prior art; in addition, it belongs to fixed working condition feature extraction and cannot achieve adaptive adjustment. It includes: when the preset cloud large model detects that the uncertainty value in any data packet is greater than the preset uncertainty threshold, the evaluation result is generated again according to the feature data in the data packet; when the evaluation result is inconsistent, the preset cloud large model feature of the preset cloud large model feature dimension corresponding to the feature data is mapped to the large model adjustment feature of the preset edge small model feature dimension; the residual between the large model adjustment feature and the corresponding preset edge small model feature is added to the preset edge small model to obtain the calibrated preset edge small model.

Description

Technical Field

[0001] This application relates to the field of adaptive detection technology, and in particular to an adaptive detection method, system, device and medium for motor sound quality based on uncertainty values. Background Technology

[0002] Current mainstream deep learning-based "small model" training relies on a fixed sample set, has a small number of model parameters, and limited signal feature extraction capabilities. It cannot adapt to dynamic changes in motor models, materials, and speeds on production lines, and the detection accuracy drops sharply when the operating conditions deviate.

[0003] Existing technologies typically employ the following methods to solve this problem: Method 1, during the training of an automatic defect detection model for industrial products, introduces a transfer learning strategy and integrates an attention mechanism based on the training image sample set to construct and optimize an industrial product defect classification model, thereby improving the model's multi-category defect recognition. Method 2, for the quality detection of surface defects in hot-rolled steel slabs, proposes a method that integrates multi-node data through federated learning and continuously updates the model using incremental learning and knowledge distillation, enhancing the model's dynamic adaptability to the production environment and thus improving the model's recognition accuracy.

[0004] However, Method 1 is time-consuming, labor-intensive, and costly, and is gradually being replaced in the era of artificial intelligence. Method 2 improves the model's generalization and recognition capabilities to some extent, but it ultimately relies on data features from other scenarios and operating conditions, which constitutes fixed-condition feature extraction. These fixed features differ from the operating conditions on the production line and cannot be adaptively adjusted, especially when the motor material, operating speed, load pressure, and fault type change. Summary of the Invention

[0005] This application provides an adaptive detection method, system, device, and medium for motor sound quality based on uncertainty values, in order to solve the problems of long training periods and high costs of existing solutions; in addition, existing solutions involve fixed-condition feature extraction and cannot achieve adaptive adjustment.

[0006] Firstly, this application provides an adaptive detection method for motor sound quality based on uncertainty values, the method comprising: Extract the training features of the preset edge small model from the training data, and adjust the training features of the preset edge small model as the feature dimensions of the preset cloud large model; based on the actual training features of the trained preset cloud large model and the feature offset of the adjusted training features, distill to obtain the trained preset edge small model. Multimodal data of the motor under operating conditions is collected, and feature data is extracted from the multimodal data. The feature data is used as input data, and K random inferences are performed through a pre-trained preset edge small model to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result. The evaluation result includes: fault defect classification and uncertainty value. Multimodal data, feature data, fault and defect classification, and uncertainty values ​​are uploaded as data packets to a pre-set cloud-based large model; at the same time, based on the fault and defect classification and uncertainty values, robotic arm operation instructions are generated and sent to the robotic arm on the production line. When the preset cloud-based large model detects that the uncertainty value in any data packet is greater than the preset uncertainty threshold, the evaluation result is generated again based on the feature data in the data packet. When the evaluation results are inconsistent, the preset cloud-based large model feature corresponding to the feature data is mapped to the large model adjustment feature of the preset edge-end small model feature dimension. The residual between the large model adjustment feature and the corresponding preset edge-end small model feature is superimposed on the preset edge-end small model to obtain the calibrated preset edge-end small model.

[0007] In one implementation of this application, preset edge-end small model training features are extracted from the training data, and these preset edge-end small model training features are mapped to adjusted training features of a preset cloud-based large model feature dimension. Based on the actual training features of the trained preset cloud-based large model and the feature offset of the adjusted training features, the trained preset edge-end small model is obtained by distillation, specifically including: Preset edge-end small model training features are extracted from the training data by preset edge-end small model; Through the formula: The preset edge small model features are the preset edge small model features of the preset edge small model feature dimensions. Mapped to adjusted training features based on preset cloud-based large model feature dimensions ; in, , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Through the formula: Calculate the feature offset; The feature offset is spatially mapped and then sent to the feature calibration module of the preset edge small model. Through the formula: The feature offset is superimposed onto the feature extraction layer output of the preset edge mini-model during the training and inference phases; in, This represents the features of the calibrated small model. Represents small model features; When the knowledge distillation of the large model is completed, the parameters of the preset edge small model are solidified, and the trained preset edge small model is obtained.

[0008] In one implementation of this application, feature data is used as input data, and K random inferences are performed through a pre-trained preset edge model to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result, specifically including: Add Dropout layers to the fully connected and convolutional layers of the preset edge small model, and use the Dropout layers to randomly deactivate a preset proportion of neurons during each forward propagation; By using a pre-defined edge-based small model to perform K random inferences on the same input feature, K predicted probability distributions are obtained. , , ,..., ; Through the formula: Calculate the uncertainty value α; in, Describes the probability distribution obtained from the i-th inference. The entropy, and C represents the total number of categories in the classification task. Let represent the probability that the input sample belongs to category c in the i-th inference; Through the formula: Calculate the integrated probability distribution; determine the highest category in the integrated probability distribution as the fault / defect classification.

[0009] In one implementation of this application, based on fault defect classification and uncertainty values, robotic arm operation instructions are generated and sent to the production line robotic arm, specifically including: When the uncertainty value is less than or equal to the preset uncertainty threshold, the robot arm operation command preset for the fault defect classification is obtained and sent to the production line robot arm. When the uncertainty value is greater than the preset uncertainty threshold, wait for the feedback result returned by the preset cloud big model; when the feedback result is consistent with the evaluation result, obtain the preset robotic arm operation instructions for the fault defect classification and send the robotic arm operation instructions to the production line robotic arm. When the feedback result is inconsistent with the evaluation result, wait for the residual data, obtain the robot arm operation instructions preset for the fault defect classification based on the fault defect classification output by the preset edge small model after residual correction, and send the robot arm operation instructions to the production line robot arm.

[0010] In one implementation of this application, the preset cloud-based large model features corresponding to the feature data are mapped to preset edge-based small model feature dimensions of the large model adjustment features, specifically including: Input the feature data into the preset cloud-based large model to obtain the preset cloud-based large model features in the feature dimensions of the preset cloud-based large model. ; Through the formula: The preset cloud-based large model features with preset cloud-based large model feature dimensions are mapped to the large model adjustment features with preset edge-based small model feature dimensions. in, Adjusting features for large models , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Features of a pre-defined cloud-based large model.

[0011] In one implementation of this application, the preset cloud-based large model features corresponding to the feature data are mapped to preset edge-based small model feature dimensions of the large model adjustment features, specifically including: Input the feature data into the preset cloud-based large model to obtain the preset cloud-based large model features in the feature dimensions of the preset cloud-based large model. ; Through the formula: The preset cloud-based large model features with preset cloud-based large model feature dimensions are mapped to the large model adjustment features with preset edge-based small model feature dimensions. in, Adjusting features for large models , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Features of a pre-defined cloud-based large model.

[0012] Secondly, this application provides an adaptive detection system for motor sound quality based on uncertainty values, the system comprising: The offline training module is used to extract preset edge small model training features from the training data, and to adjust the preset edge small model training features into preset cloud large model feature dimensions of the training features; based on the actual training features of the trained preset cloud large model and the feature offset of the adjusted training features, the trained preset edge small model is obtained by distillation. The edge module is used to collect multimodal data under motor operating conditions and extract feature data from the multimodal data. Using the feature data as input, it performs K random inferences through a pre-trained edge mini-model to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result. The evaluation result includes fault / defect classification and uncertainty value. The multimodal data, feature data, fault / defect classification, and uncertainty value are uploaded as a data package to a pre-set cloud-based large model. Simultaneously, based on the fault / defect classification and uncertainty value, robotic arm operation instructions are generated and sent to the production line robotic arm. The cloud module is used to generate an evaluation result again based on the feature data in any data packet when the preset cloud large model detects that the uncertainty value in any data packet is greater than the preset uncertainty threshold. When the evaluation results are inconsistent, the preset cloud large model feature corresponding to the feature data is mapped to the large model adjustment feature of the preset edge small model feature dimension. The residual between the large model adjustment feature and the corresponding preset edge small model feature is superimposed on the preset edge small model to obtain the calibrated preset edge small model.

[0013] In one implementation of this application, the edge module includes an acquisition unit. Used to add Dropout layers to the fully connected and convolutional layers of a predefined edge small model, and to use Dropout layers to randomly deactivate a predefined proportion of neurons during each forward propagation; By using a pre-defined edge-based small model to perform K random inferences on the same input feature, K predicted probability distributions are obtained. , , ,..., ; Through the formula: Calculate the uncertainty value α; in, Describes the probability distribution obtained from the i-th inference. The entropy, and C represents the total number of categories in the classification task. Let represent the probability that the input sample belongs to category c in the i-th inference; Through the formula: Calculate the integrated probability distribution; determine the highest category in the integrated probability distribution as the fault / defect classification.

[0014] Thirdly, this application provides an adaptive detection device for motor sound quality based on uncertainty values, the device comprising: processor; And a memory that stores executable code, which, when executed, causes the processor to perform an adaptive detection method for motor sound quality based on uncertainty values, as described above.

[0015] Fourthly, this application provides a non-volatile computer storage medium storing computer instructions, which, when executed, implement an adaptive detection method for motor sound quality based on uncertainty values, as described above.

[0016] As can be seen from the above technical solutions, this application has the following advantages: This application can accurately quantify the confidence level of a single detection result by performing K random inferences and uncertainty value calculations on edge small models, and can identify the decay nodes of the model's generalization ability without human intervention. At the same time, it adopts a collaborative mode of real-time edge inference and asynchronous cloud assistance, so that the model calibration process does not interrupt the production line operation. This avoids the problems of low production line detection reliability and the need for secondary manual verification during traditional model retraining, and solves the defects of existing technical solutions that are time-consuming, labor-intensive, have long cycles, and high application costs.

[0017] Unlike existing technologies that rely on static optimization methods that borrow data features from other scenarios, this application is based on real-time collected multimodal data of motors from the production line and actual operating conditions. Through feature space mapping between the large cloud model and the small edge model, it calculates feature offset residuals and completes feature compensation calibration. This can adapt to multi-dimensional changes in operating conditions such as motor material, operating speed, load pressure, and fault type, avoiding the feature mismatch problem between fixed features and actual operating conditions on the production line, and improving the generalization ability of the detection model in variable operating conditions.

[0018] This application achieves real-time inference and linkage with production line equipment through a lightweight small model at the edge, ensuring the low latency requirements of high-cycle detection on the production line. It eliminates the need to deploy high-computing-power equipment at the edge, reducing hardware deployment costs. At the same time, through on-demand assistance and dynamic calibration of the large model in the cloud, the strong feature extraction capabilities of the large model are empowered to the small model at the edge, continuously optimizing detection performance without affecting real-time detection on the production line. Attached Figure Description

[0019] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of an adaptive detection method for motor sound quality based on uncertainty values ​​provided in an embodiment of this application.

[0021] Figure 2 This is a block diagram illustrating an improved large-scale model knowledge distillation implementation provided in an embodiment of this application.

[0022] Figure 3 This is a schematic diagram of the internal structure of an adaptive detection system for motor sound quality based on uncertainty values, provided in an embodiment of this application.

[0023] Figure 4 This is a schematic diagram of the internal structure of an adaptive detection device for motor sound quality based on uncertainty values, provided in an embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Those skilled in the art should understand that the embodiments described below are merely preferred embodiments of this disclosure and do not imply that this disclosure can only be implemented through these preferred embodiments. These preferred embodiments are merely used to explain the technical principles of this disclosure and are not intended to limit the scope of protection of this disclosure. Based on the preferred embodiments provided by this disclosure, all other embodiments obtained by those skilled in the art without creative effort should still fall within the scope of protection of this disclosure.

[0026] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0027] This application proposes an adaptive detection method, system, device, and medium for motor acoustic quality based on uncertainty values. The core of this approach is to overcome the shortcomings of existing technologies, such as insufficient generalization ability of small models, poor adaptability to changing operating conditions, and low detection accuracy, through adaptive adjustment. The system operates within a cloud-edge-device architecture. A large acoustic quality detection model is deployed in the cloud, connecting with a smaller model on the production line. The large model provides knowledge transfer and feature calibration for the smaller model, addressing the issue of decreased detection accuracy due to changes in motor material, fluctuating operating conditions, and environmental interference. The smaller model is deployed on the production line, detecting the acoustic quality characteristics of each motor in real time, recording the detection data, and uploading it to the enterprise's QMS (Quality Management System) to meet subsequent quality traceability requirements. This adaptive motor acoustic quality detection system, combining a large model and a small model, adopts a collaborative working mode of "real-time edge inference + dynamic cloud assistance," solving the current problems in motor acoustic quality detection and achieving a balance between high accuracy, high adaptability, low latency, and low cost in production line motor acoustic quality detection.

[0028] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0029] The embodiment provides an adaptive detection method for motor sound quality based on uncertainty values, such as Figure 1 As shown in the embodiments of this application, the method mainly includes the following steps: Step 110: Extract the preset edge small model training features from the training data, and adjust the preset edge small model training features as the feature dimensions of the preset cloud large model; based on the actual training features of the trained preset cloud large model and the feature offset of the adjusted training features, distill to obtain the trained preset edge small model.

[0030] It should be noted that this specifically includes: Preset edge-end small model training features are extracted from the training data by preset edge-end small model; Through the formula: The preset edge small model features are the preset edge small model features of the preset edge small model feature dimensions. Mapped to adjusted training features based on preset cloud-based large model feature dimensions ; in, , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Through the formula: Calculate the feature offset; The feature offset is spatially mapped and then sent to the feature calibration module of the preset edge small model. Through the formula: The feature offset is superimposed onto the feature extraction layer output of the preset edge mini-model during the training and inference phases; in, This represents the features of the calibrated small model. Represents small model features; When the knowledge distillation of the large model is completed, the parameters of the preset edge small model are solidified, and the trained preset edge small model is obtained.

[0031] Step 120: Collect multimodal data under motor operating conditions, extract feature data from the multimodal data; use the feature data as input data, perform K random inferences through the trained preset edge small model to obtain K predicted probability distributions; sum and average the entropy of the K predicted probability distributions to obtain the evaluation result.

[0032] The evaluation results include: fault / defect classification and uncertainty value.

[0033] In some embodiments, feature data is used as input data, and K random inferences are performed through a pre-trained preset edge model to obtain K predicted probability distributions; the entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result, specifically including: Add Dropout layers to the fully connected and convolutional layers of the preset edge small model, and use the Dropout layers to randomly deactivate a preset proportion of neurons during each forward propagation; By using a pre-defined edge-based small model to perform K random inferences on the same input feature, K predicted probability distributions are obtained. , , ,..., ; Through the formula: Calculate the uncertainty value α; in, Describes the probability distribution obtained from the i-th inference. The entropy, and C represents the total number of categories in the classification task. Let represent the probability that the input sample belongs to category c in the i-th inference; Through the formula: Calculate the integrated probability distribution; determine the highest category in the integrated probability distribution as the fault / defect classification.

[0034] Understandably, by introducing Dropout layers into the fully connected and convolutional layers of the edge small model, a predetermined proportion of neurons are randomly deactivated during each forward propagation, enabling the model to perform K random inferences on the same input feature, generating diverse probability distributions. By summing and averaging the entropy of each inference result, a quantified uncertainty value is obtained. This value directly reflects the model's confidence in the current input and is generated synchronously with the fault / defect classification results. No additional computation module is required, enabling the simultaneous output of classification labels and confidence assessments within a single inference process.

[0035] In addition, when the operating conditions of the motor deviate slightly (such as load fluctuations or slight bearing wear), traditional deterministic models may output high-confidence misclassifications. This solution identifies low-confidence samples through an entropy averaging mechanism, so that the system maintains the original classification output when the uncertainty value does not exceed the threshold, avoiding unnecessary robotic arm intervention or alarms triggered by model misjudgment, and reducing the risk of unnecessary production line downtime.

[0036] Step 130: Upload the multimodal data, feature data, fault and defect classification, and uncertainty value as a data package to the preset cloud-based large model; at the same time, generate robotic arm operation instructions based on the fault and defect classification and uncertainty value, and send the robotic arm operation instructions to the production line robotic arm.

[0037] In some embodiments, based on fault defect classification and uncertainty values, robotic arm operation instructions are generated and sent to the production line robotic arm, specifically including: When the uncertainty value is less than or equal to the preset uncertainty threshold, the robot arm operation command is obtained according to the preset fault defect classification (e.g., qualified, rotor imbalance, bearing noise, vibration, abnormal noise, etc.) and sent to the production line robot arm. When the uncertainty value is greater than the preset uncertainty threshold, wait for the feedback result returned by the preset cloud big model; when the feedback result is consistent with the evaluation result, obtain the preset robotic arm operation instructions for the fault defect classification and send the robotic arm operation instructions to the production line robotic arm. When the feedback result is inconsistent with the evaluation result, wait for the residual data, obtain the robot arm operation instructions preset for the fault defect classification based on the fault defect classification output by the preset edge small model after residual correction, and send the robot arm operation instructions to the production line robot arm.

[0038] It should be noted that a preset uncertainty threshold τ (τ = 0.3~0.5, which can be adjusted according to production line requirements) is set: when α ≤ τ, it indicates that the confidence of the small model inference is high and the sample is within the reliable generalization range of the small model, and the inference result is directly output. When α>τ, it indicates that the current sample exceeds the reliable generalization range of the small model (possibly due to changes in material, state fluctuations, etc.), and is marked as a "difficult sample".

[0039] Step 140: When the preset cloud-based large model detects that the uncertainty value in any data packet is greater than the preset uncertainty threshold, the evaluation result is generated again based on the feature data in the data packet; when the evaluation results are inconsistent, the preset cloud-based large model feature corresponding to the feature data is mapped to the large model adjustment feature of the preset edge-end small model feature dimension; the residual between the large model adjustment feature and the corresponding preset edge-end small model feature is superimposed on the preset edge-end small model to obtain the calibrated preset edge-end small model.

[0040] It should be noted that the large model sets an uncertainty threshold τ (τ = 0.3~0.5, which can be adjusted according to production line requirements) based on the differences in production lines: when α ≤ τ, it indicates that the small model has high inference confidence and the sample is within the reliable generalization range of the small model, and the inference result is directly output. When α > τ, it indicates that the current sample exceeds the reliable generalization range of the small model (possibly due to changes in material, state fluctuations, etc.), and is marked as a "difficult sample". In order not to affect the working efficiency of the small model, data transmission can adopt an asynchronous working mode and supports packaged and compressed transmission.

[0041] A pre-defined cloud-based large model performs feature extraction and forward computation on the uploaded, problematic sample raw signals to obtain inference results. If the large model's inference result matches the small model's inference result, the small model's inference is correct and no adjustment is needed. If the large model's inference result does not match the small model's inference result, the small model's inference result is incorrect, and the small model's parameters need to be adjusted based on the large model's inference result to ensure the small model's generalization ability for subsequent similar operating conditions.

[0042] In some embodiments, mapping the preset cloud-based large model features of the preset cloud-based large model feature dimension corresponding to the feature data to the large model adjustment features of the preset edge-based small model feature dimension specifically includes: Input the feature data into the preset cloud-based large model to obtain the preset cloud-based large model features in the feature dimensions of the preset cloud-based large model. ; Through the formula: The preset cloud-based large model features with preset cloud-based large model feature dimensions are mapped to the large model adjustment features with preset edge-based small model feature dimensions. in, Adjusting features for large models , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Features of a pre-defined cloud-based large model.

[0043] In some embodiments, mapping the preset cloud-based large model features of the preset cloud-based large model feature dimension corresponding to the feature data to the large model adjustment features of the preset edge-based small model feature dimension specifically includes: Input the feature data into the preset cloud-based large model to obtain the preset cloud-based large model features in the feature dimensions of the preset cloud-based large model. ; Through the formula: The preset cloud-based large model features with preset cloud-based large model feature dimensions are mapped to the large model adjustment features with preset edge-based small model feature dimensions. in, Adjusting features for large models , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Features of a pre-defined cloud-based large model.

[0044] It should be noted that a large model dataset was built based on accumulated historical data. This dataset contains data from various motor materials, environments, and operating conditions, covering both compliant and non-compliant samples, and contains more than 100,000 samples. This dataset has been continuously refined and iterated since the system went live.

[0045] Based on the preceding description, the technical implementation process is shown in Figure 2. Unlike conventional knowledge distillation implementation schemes, this application improves the large-model knowledge distillation technology by adding feature space mapping and feature calibration modules. Feature space mapping maps low-dimensional features of the small model to high-dimensional feature spaces of the large model, satisfying feature alignment and facilitating the large model's ability to identify operational anomalies in the small model under subsequent variable operating conditions. The feature fusion module, based on the large model's acquisition of production line sensing data, accesses the MES system to obtain parameters related to the current production, such as product model, production batch, product material, and environmental noise. After vectorization processing, these parameters are used as input features for the large model's inference through feature concatenation.

[0046] As described above, this embodiment achieves significant technical and business value in motor sound quality detection scenarios through a collaborative mechanism between large and small models. In terms of technical effectiveness, the system can perceive changes in operating conditions in real time. When batches of motor materials are changed, operating states fluctuate, or environmental noise interference causes drift in detection features, the cloud-based large model automatically identifies abnormal trends and sends calibration parameters. This allows the production line small model to quickly restore detection accuracy without retraining, effectively solving the problems of poor generalization ability and insufficient adaptability to changing operating conditions inherent in traditional small models. This significantly reduces the false alarm rate and false negative rate.

[0047] In terms of business value, this invention supports the high-cycle, real-time inspection needs of production lines, with the inspection latency of a single motor controllable at the millisecond level, meeting the response speed requirements of large-scale production. Simultaneously, the inspection results are automatically uploaded to the enterprise quality management system, building a complete sound quality inspection file for each motor, supporting quality traceability, process analysis, and batch quality assessment. This system, through lightweight edge deployment combined with cloud-based intelligent assistance, avoids frequent model changes or reliance on high-cost computing equipment. While ensuring inspection accuracy, it achieves an effective balance between deployment costs, maintenance costs, and inspection performance, possessing good versatility and scalability, and can be widely applied to online quality inspection scenarios for motors and similar rotating machinery.

[0048] In addition, this application Figure 3 This application provides an embodiment of an adaptive detection system for motor sound quality based on uncertainty values. For example... Figure 3 As shown in the embodiments of this application, the system mainly includes: The offline training module 210 is used to extract preset edge-end small model training features from the training data, and to adjust the preset edge-end small model training features into preset cloud-based large model feature dimensions; based on the actual training features of the trained preset cloud-based large model and the feature offset of the adjusted training features, the trained preset edge-end small model is obtained by distillation. Edge module 220 is used to collect multimodal data under motor operating conditions and extract feature data from the multimodal data. The feature data is used as input data and K random inferences are performed through a pre-trained preset edge small model to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result. The evaluation result includes fault defect classification and uncertainty value. The multimodal data, feature data, fault defect classification and uncertainty value are uploaded as a data package to a preset cloud large model. At the same time, based on the fault defect classification and uncertainty value, robotic arm operation instructions are generated and sent to the production line robotic arm.

[0049] Edge module 220 includes an acquisition unit. Used to add Dropout layers to the fully connected and convolutional layers of a predefined edge small model, and to use Dropout layers to randomly deactivate a predefined proportion of neurons during each forward propagation; By using a pre-defined edge-based small model to perform K random inferences on the same input feature, K predicted probability distributions are obtained. , , ,..., ; Through the formula: Calculate the uncertainty value α; in, Describes the probability distribution obtained from the i-th inference. The entropy, and C represents the total number of categories in the classification task. Let represent the probability that the input sample belongs to category c in the i-th inference; Through the formula: Calculate the integrated probability distribution; determine the highest category in the integrated probability distribution as the fault / defect classification.

[0050] The cloud module 230 is used to generate an evaluation result again based on the feature data in the data packet when the preset cloud large model detects that the uncertainty value in any data packet is greater than the preset uncertainty threshold; when the evaluation results are inconsistent, the preset cloud large model feature corresponding to the feature data is mapped to the large model adjustment feature of the preset edge small model feature dimension; the residual between the large model adjustment feature and the corresponding preset edge small model feature is superimposed on the preset edge small model to obtain the calibrated preset edge small model.

[0051] The above are method embodiments of this application. Based on the same inventive concept, this application also provides an adaptive detection device for motor sound quality based on uncertainty values. Figure 4 As shown, the device includes: a processor; and a memory storing executable code thereon, which, when executed, causes the processor to perform an adaptive detection method for motor sound quality based on uncertainty values ​​as described in the above embodiments.

[0052] Specifically, the server extracts pre-defined edge-end small model training features from the training data, and adjusts these features to reflect the feature dimensions of the pre-defined cloud-based large model. Based on the actual training features of the pre-trained pre-defined cloud-based large model and the feature offsets of the adjusted training features, distillation is performed to obtain the pre-trained pre-defined edge-end small model. Multimodal data under motor operating conditions is collected, and feature data is extracted from the multimodal data. Using the feature data as input data, the pre-trained pre-defined edge-end small model is used to perform K random inferences to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result. The evaluation result includes: fault / defect classification and uncertainty value. The multimodal data and features are then combined. Data, fault / defect classifications, and uncertainty values ​​are uploaded as data packets to a pre-defined cloud-based large model. Simultaneously, based on the fault / defect classifications and uncertainty values, robotic arm operation instructions are generated and sent to the production line robotic arm. When the pre-defined cloud-based large model detects that the uncertainty value in any data packet exceeds a pre-defined uncertainty threshold, it regenerates an evaluation result based on the feature data in the data packet. When the evaluation results are inconsistent, the pre-defined cloud-based large model features corresponding to the feature data are mapped to the large model adjustment features of the pre-defined edge-end small model features. The residual between the large model adjustment features and the corresponding pre-defined edge-end small model features is superimposed onto the pre-defined edge-end small model to obtain the calibrated pre-defined edge-end small model.

[0053] In addition, embodiments of this application also provide a non-volatile computer storage medium storing executable instructions, which, when executed, implement the aforementioned adaptive detection method for motor sound quality based on uncertainty values.

[0054] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An adaptive detection method for motor sound quality based on uncertainty values, characterized in that, The method includes: Extract the training features of the preset edge small model from the training data, and adjust the training features of the preset edge small model as the feature dimensions of the preset cloud large model; based on the actual training features of the trained preset cloud large model and the feature offset of the adjusted training features, distill to obtain the trained preset edge small model. Multimodal data of the motor under operating conditions is collected, and feature data is extracted from the multimodal data. The feature data is used as input data, and K random inferences are performed through a pre-trained preset edge small model to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result. The evaluation result includes: fault defect classification and uncertainty value. Multimodal data, feature data, fault and defect classification, and uncertainty values ​​are uploaded as data packets to a pre-set cloud-based large model; at the same time, based on the fault and defect classification and uncertainty values, robotic arm operation instructions are generated and sent to the robotic arm on the production line. When the preset cloud-based large model detects that the uncertainty value in any data packet is greater than the preset uncertainty threshold, the evaluation result is generated again based on the feature data in the data packet. When the evaluation results are inconsistent, the preset cloud-based large model feature corresponding to the feature data is mapped to the large model adjustment feature of the preset edge-end small model feature dimension. The residual between the large model adjustment feature and the corresponding preset edge-end small model feature is superimposed on the preset edge-end small model to obtain the calibrated preset edge-end small model.

2. The adaptive detection method for motor sound quality based on uncertainty value according to claim 1, characterized in that, Extract preset edge small model training features from the training data, and map the preset edge small model training features to preset cloud large model feature dimension adjustment training features; Based on the actual training features of the pre-trained preset cloud-based large model and the feature offsets of the adjusted training features, distillation is performed to obtain the pre-trained preset edge-based small model, specifically including: Preset edge-end small model training features are extracted from the training data by preset edge-end small model; Through the formula: The preset edge small model features are the preset edge small model features of the preset edge small model feature dimensions. Mapped to adjusted training features based on preset cloud-based large model feature dimensions ; in, , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Through the formula: Calculate the feature offset; The feature offset is spatially mapped and then sent to the feature calibration module of the preset edge small model. Through the formula: The feature offset is superimposed onto the feature extraction layer output of the preset edge mini-model during the training and inference phases; in, This represents the features of the calibrated small model. Represents small model features; When the knowledge distillation of the large model is completed, the parameters of the preset edge small model are solidified, and the trained preset edge small model is obtained.

3. The adaptive detection method for motor sound quality based on uncertainty values ​​according to claim 1, characterized in that, Using feature data as input, a pre-trained edge-end mini-model performs K random inferences to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result, which specifically includes: Add Dropout layers to the fully connected and convolutional layers of the preset edge small model, and use the Dropout layers to randomly deactivate a preset proportion of neurons during each forward propagation; By using a pre-defined edge-based small model to perform K random inferences on the same input feature, K predicted probability distributions are obtained. , , ,..., ; Through the formula: Calculate the uncertainty value α; in, Describes the probability distribution obtained from the i-th inference. The entropy, and C represents the total number of categories in the classification task. Let represent the probability that the input sample belongs to category c in the i-th inference; Through the formula: Calculate the integrated probability distribution; determine the highest category in the integrated probability distribution as the fault / defect classification.

4. The adaptive detection method for motor sound quality based on uncertainty value according to claim 1, characterized in that, Based on fault and defect classification and uncertainty values, robotic arm operation instructions are generated and sent to the production line robotic arm. These instructions specifically include: When the uncertainty value is less than or equal to the preset uncertainty threshold, the robot arm operation command preset for the fault defect classification is obtained and sent to the production line robot arm. When the uncertainty value is greater than the preset uncertainty threshold, wait for the feedback result returned by the preset cloud big model; when the feedback result is consistent with the evaluation result, obtain the preset robotic arm operation instructions for the fault defect classification and send the robotic arm operation instructions to the production line robotic arm. When the feedback result is inconsistent with the evaluation result, wait for the residual data, obtain the robot arm operation instructions preset for the fault defect classification based on the fault defect classification output by the preset edge small model after residual correction, and send the robot arm operation instructions to the production line robot arm.

5. The adaptive detection method for motor sound quality based on uncertainty values ​​according to claim 1, characterized in that, Mapping the preset cloud-based large model features corresponding to the feature data to the preset edge-based small model feature dimensions of the large model adjustment features, specifically including: Input the feature data into the preset cloud-based large model to obtain the preset cloud-based large model features in the feature dimensions of the preset cloud-based large model. ; Through the formula: The preset cloud-based large model features with preset cloud-based large model feature dimensions are mapped to the large model adjustment features with preset edge-based small model feature dimensions. in, Adjusting features for large models , These are the weight matrices for the first and second layers of a fully connected neural network, respectively. , These are the bias vectors for the first and second layers of the fully connected neural network, respectively. For activation functions; Features of a pre-defined cloud-based large model.

6. The adaptive detection method for motor sound quality based on uncertainty value according to claim 1, characterized in that, The residuals between the large model adjustment features and the corresponding preset edge-small model features are superimposed onto the preset edge-small model to obtain the calibrated preset edge-small model, specifically including: Input the feature data into the trained preset edge mini-model to obtain the preset edge mini-model features in the preset edge mini-model feature dimension. ; Through the formula: Calculate large model adjustment features and corresponding preset edge small model features The residuals between; Through the formula: The residuals are superimposed onto the preset edge small model to obtain the calibrated preset edge small model. in, Indicates the calibration coefficient. This represents the original preset edge small model features.

7. An adaptive detection system for motor sound quality based on uncertainty values, characterized in that, The system includes: The offline training module is used to extract preset edge small model training features from the training data, and to adjust the preset edge small model training features into preset cloud large model feature dimensions of the training features; based on the actual training features of the trained preset cloud large model and the feature offset of the adjusted training features, the trained preset edge small model is obtained by distillation. The edge module is used to collect multimodal data under motor operating conditions and extract feature data from the multimodal data. Using the feature data as input, it performs K random inferences through a pre-trained edge mini-model to obtain K predicted probability distributions. The entropy of the K predicted probability distributions is summed and averaged to obtain the evaluation result. The evaluation result includes fault / defect classification and uncertainty value. The multimodal data, feature data, fault / defect classification, and uncertainty value are uploaded as a data package to a pre-set cloud-based large model. Simultaneously, based on the fault / defect classification and uncertainty value, robotic arm operation instructions are generated and sent to the production line robotic arm. The cloud module is used to generate an evaluation result again based on the feature data in any data packet when the preset cloud large model detects that the uncertainty value in any data packet is greater than the preset uncertainty threshold. When the evaluation results are inconsistent, the preset cloud large model feature corresponding to the feature data is mapped to the large model adjustment feature of the preset edge small model feature dimension. The residual between the large model adjustment feature and the corresponding preset edge small model feature is superimposed on the preset edge small model to obtain the calibrated preset edge small model.

8. The adaptive detection system for motor sound quality based on uncertainty value according to claim 7, characterized in that, The edge module includes an acquisition unit. Used to add Dropout layers to the fully connected and convolutional layers of a predefined edge small model, and to use Dropout layers to randomly deactivate a predefined proportion of neurons during each forward propagation; By using a pre-defined edge-based small model to perform K random inferences on the same input feature, K predicted probability distributions are obtained. , , ,..., ; Through the formula: Calculate the uncertainty value α; in, Describes the probability distribution obtained from the i-th inference. The entropy, and C represents the total number of categories in the classification task. Let represent the probability that the input sample belongs to category c in the i-th inference; Through the formula: Calculate the integrated probability distribution; determine the highest category in the integrated probability distribution as the fault / defect classification.

9. An adaptive detection device for motor sound quality based on uncertainty values, characterized in that, The device includes: processor; And a memory storing executable code, which, when executed, causes the processor to perform an adaptive detection method for motor sound quality based on uncertainty values ​​as described in any one of claims 1-6.

10. A non-volatile computer storage medium, characterized in that, It stores computer instructions, which, when executed, implement an adaptive detection method for motor sound quality based on uncertainty values ​​as described in any one of claims 1-6.

Images