Intelligent tightening tool capacity deployment system and method based on deep learning

Abstract

The application discloses an intelligent tightening tool production capacity allocation system and method based on deep learning, relates to the technical field of tightening tools, and has the advantages that multi-modal data is captured in real time, data weights are dynamically adjusted by using an attention mechanism, production capacity is predicted by combining a deep learning algorithm, dynamic production requirements and emergent tasks are quickly responded to, a variational autoencoder VAE model is introduced in the aspect of abnormality processing, double indexes of reconstruction error and distribution deviation are used to capture abnormal modes, a probability distribution model of normal operation data is established by combining unsupervised learning, false positive and false negative rates are significantly reduced, abnormal data is fed back to a production capacity prediction module in real time through a closed loop mechanism, deep learning model parameters are dynamically optimized, the system can adapt to changes in production tasks and attenuation of tool performance, and a deployment module is used to allocate tool types and optimize paths according to task priorities, so that resource utilization efficiency is improved, tool conflicts and workstation waiting time are reduced.

Description

Technical Field

[0001] This invention relates to the field of tightening tool technology, and in particular to a deep learning-based intelligent tightening tool capacity allocation system and method. Background Technology

[0002] In modern industrial production, tightening tools are an important component of the assembly line, and the quality of their capacity allocation directly affects overall production efficiency and product quality. In recent years, some allocation schemes have introduced automation and intelligent technologies, but their responsiveness to dynamic production demands remains insufficient. In the assembly line environment, production task requirements frequently change, such as new tasks, priority adjustments, or equipment failures. However, most current allocation strategies are still based on static rules or periodic calculations, making it difficult to capture and respond to real-time changes in a timely manner. This lag often leads to task queuing, resource idleness, or unbalanced capacity allocation, thus adversely affecting production efficiency.

[0003] To address the issue of delayed response in allocation, some traditional solutions employ manual intervention or fixed rules. However, these methods have limitations such as slow response speed and limited rules, making it difficult to meet the demands of high-frequency adjustments. Therefore, there is an urgent need for a deep learning-based intelligent tightening tool capacity allocation system and method to solve this problem. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] This invention provides a deep learning-based intelligent tightening tool capacity allocation system and method to solve the problems of slow allocation response in traditional solutions, and the limitations of some solutions that rely on manual intervention or fixed rules, such as slow response speed and single rules, which make it difficult to meet the needs of high-frequency adjustment.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a deep learning-based intelligent tightening tool capacity allocation system, which includes a data acquisition module, a data upload module, a capacity prediction module, an allocation module, an error warning module, a feedback module, and a user interface.

[0008] The data acquisition module has a built-in controller, which uses PLC and PC-based controllers to collect the working data of the tightening tool, including sensor data, production history and worker operation records;

[0009] Sensor data includes torque, speed, and temperature and humidity;

[0010] Production history includes task sequence and tool usage time;

[0011] Worker operation records include operation frequency and real-time intervention data;

[0012] The data upload module is used to upload working data to the capacity forecasting module;

[0013] The capacity prediction module uses deep learning algorithms to predict capacity based on working data and adjusts production strategies accordingly.

[0014] The allocation module adjusts production capacity according to the production strategy, including the type and quantity of tools at each workstation, as well as the tightening sequence and tightening path;

[0015] The error warning module is used to analyze abnormal data during the tightening process, including abnormal torque and abnormal speed, and to provide error warnings.

[0016] The feedback module is used to feed back abnormal data to the capacity prediction module in real time;

[0017] The user interface is used by operators to monitor and adjust the status of tightening tools.

[0018] As a preferred embodiment of the intelligent tightening tool capacity allocation system based on deep learning described in this invention, the data acquisition module adopts a multimodal fusion model based on an attention mechanism to dynamically adjust the weight of the work data, generate comprehensive work features, and input the fused work features into the allocation module and the error warning module.

[0019] As a preferred embodiment of the intelligent tightening tool capacity allocation system based on deep learning described in this invention, the error warning module constructs an anomaly detection model based on variational autoencoder (VAE) with the fused working features as input; it adopts unsupervised learning to establish a probability distribution model of normal operation data, and during operation, it uses the reconstruction error and distribution deviation of real-time data as anomaly indicators, and combines business rules to perform multi-level warnings.

[0020] Secondly, this invention provides a deep learning-based intelligent tightening tool capacity allocation method, comprising:

[0021] Step S1: Collect working data during the tightening process using the data acquisition module.

[0022] Step S2 involves using a multimodal fusion model based on an attention mechanism to dynamically adjust the weights of the collected work data, generating comprehensive work features. These fused work features are then input into the allocation module and the error warning module.

[0023] Step S3: The capacity prediction module uses a deep learning algorithm to predict capacity based on the fused working characteristics and adjusts production strategies accordingly.

[0024] In step S4, the allocation module adjusts production capacity according to the production strategy, including the tool type and quantity at each workstation, as well as the tightening sequence and tightening path; at the same time, the error warning module constructs an anomaly detection model based on a variational autoencoder (VAE), using the fused work features as input to analyze abnormal data in the tightening process, including torque anomalies and speed anomalies; and uses unsupervised learning to establish a probability distribution model of normal operation data. During operation, the reconstruction error and distribution deviation of real-time data are used as anomaly indicators, and multi-level warnings are issued in conjunction with business rules.

[0025] In step S5, the feedback module sends the abnormal data back to the capacity prediction module to update the deep learning model in step S3.

[0026] As a preferred embodiment of the deep learning-based intelligent tightening tool capacity allocation method described in this invention, the step of dynamically adjusting the weights of the collected work data using a multimodal fusion model based on an attention mechanism to generate comprehensive work features is as follows:

[0027] Data preprocessing is performed, and after standardization of the multimodal data, preliminary features are extracted using the following formula:

[0028] ,

[0029] in, The feature matrix of the sensor data, The feature matrix of production history. The feature matrix of worker operation records. For the raw data input of the corresponding mode, These are convolutional neural networks and recurrent neural networks, respectively.

[0030] The weights of each modal feature are dynamically adjusted based on an attention mechanism, and the calculation formula is as follows:

[0031] ,

[0032] in, For modality The weight, , For modality Preliminary characteristics The learnable parameter matrix used for weight calculation, This is a weighted normalized modal index, with a value range of [value range missing]. ,

[0033] After fusing modal features according to weights, the final comprehensive feature is generated. The formula for the comprehensive feature is:

[0034] ,

[0035] in, To comprehensively consider the characteristics of the work, This is the weight matrix of the fully connected layer. For the bias of the fully connected layer, It is a non-linear activation function; the attention mechanism dynamically adjusts the modality weights to effectively generate high-quality integrated working features.

[0036] As a preferred embodiment of the deep learning-based intelligent tightening tool capacity allocation method described in this invention, the steps of predicting capacity and allocating production strategies based on fused working characteristics using a deep learning algorithm are as follows:

[0037] The capacity prediction module uses a deep feedforward neural network (DNN) for modeling, and the calculation process is as follows:

[0038] ,

[0039] ,

[0040] in, For the first The output features of the layer The overall working characteristics of the input layer, For the first Layer weight matrix and bias, This is the final capacity forecast. The total number of layers in the network;

[0041] Using the mean squared error (MSE) as the loss function, it is defined as follows: :

[0042] ,

[0043] The model parameters are then updated using gradient descent, with the following update formula:

[0044] ,

[0045] in, The loss value represents the mean squared error between the predicted and actual values. For the sample size, For the first The predicted value for each sample, For the first The actual production capacity of each sample Network parameters, including weights and biases. For learning rate, The number of iterations is used here. A deep learning model is employed to capture data patterns, thereby improving the accuracy of capacity forecasting.

[0046] As a preferred embodiment of the intelligent tightening tool capacity allocation method based on deep learning described in this invention, the step of the allocation module adjusting capacity according to the production strategy is as follows:

[0047] The allocation module adjusts production capacity, including:

[0048] According to the task priority allocation tool, the allocation formula is as follows:

[0049] ,

[0050] in, To be assigned to workstations Tool types, For tools At workstation The fit score on the screen

[0051] The path optimization formula is as follows:

[0052] ,

[0053] in, The total cost of the path. For path distance, For path Task weights For path The allocation module improves capacity utilization through intelligent allocation and path optimization, while the error warning module ensures production stability through anomaly detection.

[0054] As a preferred embodiment of the deep learning-based intelligent tightening tool capacity allocation method described in this invention, the steps of constructing an anomaly detection model based on a variational autoencoder (VAE), and analyzing abnormal data during the tightening process using fused working features as input, are as follows:

[0055] The latent distribution is constructed using a variational autoencoder (VAE), and the construction formula is as follows:

[0056] ,

[0057] in, Representation of the potential space, The mean and standard deviation of the potential distribution are given. It is random noise. Features for reconstruction.

[0058] As a preferred embodiment of the deep learning-based intelligent tightening tool capacity allocation method described in this invention, the step of using the reconstruction error and distribution deviation of real-time data as anomaly indicators and combining them with business rules for multi-level early warning is as follows:

[0059] The reconstruction error is calculated using the following formula:

[0060] ,

[0061] in, Reconstruction error, representing the input features. Reconstructed output with autoencoder The differences between them To synthesize the overall working characteristics, the input feature matrix is ​​fused by the attention mechanism. The synthesized features reconstructed by the autoencoder are then processed by the decoder. From potential representation The generated feature matrix, The square of the L2 norm;

[0062] The distribution bias is calculated using the following formula:

[0063] ,

[0064] in, Let be the distribution bias, representing the approximate posterior distribution of the latent space. With prior distribution Kullback-Leibler divergence between them It is an approximate posterior distribution, based on input features. The generated latent distribution is a Gaussian distribution: , The mean vector is an approximate posterior distribution, generated by the encoder. The standard deviation vector of the approximate posterior distribution is generated by the encoder. The prior distribution is , and the standard normal distribution is . , Let be the mean vector of the prior distribution. Let KL be the identity matrix, representing the independent and identically distributed standard normal distribution. The Kullback-Leibler divergence, used to measure the relative entropy difference between two probability distributions, is defined as:

[0065] ,

[0066] By calculating the reconstruction error and distribution deviation The system evaluates whether the input features conform to the normal operating state of the model. If the deviation is too large, it indicates an anomaly. Anomaly indicators together constitute the basis for anomaly detection in the autoencoder.

[0067] As a preferred embodiment of the intelligent tightening tool capacity allocation method based on deep learning described in this invention, the feedback module compares abnormal data with prediction results to generate optimized samples. The optimized samples are used to update the deep learning model. The feedback module forms a closed-loop system, which enhances the model's ability to handle complex scenarios.

[0068] The beneficial effects of this invention are as follows: This invention captures multimodal data in real time, dynamically adjusts data weights using an attention mechanism, and combines deep learning algorithms to predict production capacity, enabling rapid response to dynamic production demands and unexpected tasks; in terms of anomaly handling, it introduces a variational autoencoder (VAE) model to reconstruct both error and distribution bias indicators, captures abnormal patterns, and combines unsupervised learning to establish a probability distribution model of normal operation data, significantly reducing false alarms and false negatives and enhancing the robustness of the system; the feedback module feeds abnormal data back to the production capacity prediction module in real time through a closed-loop mechanism, dynamically optimizing the parameters of the deep learning model, enabling the system to adapt to changes in production tasks and the degradation of tool performance, achieving long-term self-learning and continuous optimization.

[0069] This invention employs a scheduling module to optimize tool types and paths by prioritizing tasks, thereby improving resource utilization efficiency and reducing tool conflicts and workstation waiting time. It also performs multimodal data fusion to reduce the impact of single-modal data anomalies on system performance, improving overall stability and decision-making accuracy. Furthermore, it combines a multi-level early warning mechanism to proactively warn of anomalies before they occur, ensuring the safety of the production process and the consistency of product quality. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0071] Figure 1 This is a schematic diagram of the capacity allocation and error warning process of the intelligent tightening tool capacity allocation system of the present invention;

[0072] Figure 2 This is a schematic diagram of the intelligent tightening tool capacity allocation system of the present invention;

[0073] Figure 3 This is a flowchart of the allocation module of the intelligent tightening tool capacity allocation system of the present invention;

[0074] Figure 4 This is a flowchart illustrating the error warning module of the intelligent tightening tool capacity allocation system of the present invention. Detailed Implementation

[0075] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0076] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0077] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0078] Example 1, referring to Figures 1-4 This embodiment provides a deep learning-based intelligent tightening tool production capacity allocation system, including:

[0079] The system includes a data acquisition module, a data upload module, a capacity forecasting module, a dispatching module, an error warning module, a feedback module, and a user interface.

[0080] The data acquisition module has a built-in controller, using PLC and PC-based controllers, to collect the working data of the tightening tool, including sensor data, production history and worker operation records;

[0081] Sensor data includes torque, speed, and temperature and humidity;

[0082] Production history includes task sequence and tool usage time;

[0083] Worker operation records include operation frequency and real-time intervention data;

[0084] In the data acquisition module, a multimodal fusion model based on the attention mechanism is adopted to dynamically adjust the weight of the work data and generate comprehensive work features. The fused work features are then input into the allocation module and the error warning module.

[0085] The data upload module is used to upload working data to the capacity forecasting module;

[0086] The capacity forecasting module uses deep learning algorithms based on work data to predict capacity and adjust production strategies accordingly.

[0087] The allocation module adjusts production capacity according to the production strategy, including the type and quantity of tools at each workstation, as well as the tightening sequence and tightening path;

[0088] The error warning module is used to analyze abnormal data during the tightening process, including abnormal torque and speed, and provide error warnings.

[0089] In the error warning module, an anomaly detection model based on variational autoencoder (VAE) is constructed, with the fused working features as input; an unsupervised learning model is used to establish the probability distribution model of normal operation data. During operation, the reconstruction error and distribution deviation of real-time data are used as anomaly indicators, and multi-level warnings are performed in combination with business rules.

[0090] The feedback module is used to feed back abnormal data to the capacity forecasting module in real time;

[0091] A user interface for operators to monitor and adjust the status of tightening tools.

[0092] This embodiment also provides a deep learning-based intelligent tightening tool capacity allocation method, including:

[0093] Step S1: Collect working data during the tightening process using the data acquisition module.

[0094] Step S2 involves using a multimodal fusion model based on an attention mechanism to dynamically adjust the weights of the collected work data, generating comprehensive work features. These fused work features are then input into the allocation module and the error warning module.

[0095] The steps for generating comprehensive work features by using an attention-based multimodal fusion model to dynamically adjust the weights of the collected work data are as follows:

[0096] Data preprocessing is performed, and after standardization of the multimodal data, preliminary features are extracted using the following formula:

[0097] ,

[0098] in, The feature matrix of the sensor data, The feature matrix of production history. The feature matrix of worker operation records. For the raw data input of the corresponding mode, These are convolutional neural networks and recurrent neural networks, respectively.

[0099] The weights of each modal feature are dynamically adjusted based on an attention mechanism, and the calculation formula is as follows:

[0100] ,

[0101] in, For modality The weight, , For modality Preliminary characteristics The learnable parameter matrix used for weight calculation, This is a weighted normalized modal index, with a value range of [value range missing]. ,

[0102] After fusing modal features according to weights, the final comprehensive feature is generated. The formula for the comprehensive feature is:

[0103] ,

[0104] in, To comprehensively consider the characteristics of the work, This is the weight matrix of the fully connected layer. For the bias of the fully connected layer, It is a non-linear activation function; the attention mechanism dynamically adjusts the modality weights to effectively generate high-quality integrated working features;

[0105] Specifically, the data acquisition module and data upload module capture multimodal data in real time during the tightening process, and combine them with an attention mechanism to dynamically adjust the weights and generate comprehensive working characteristics. On this basis, the capacity prediction module uses a deep feedforward neural network (DNN) to model and quickly predict future capacity, generating dynamic strategies. The allocation module uses the prediction results to quickly adjust task allocation and path optimization, which significantly improves the real-time response capability of the production line.

[0106] Step S3: The capacity prediction module uses a deep learning algorithm to predict capacity based on the fused working characteristics and adjusts production strategies accordingly.

[0107] Based on the fused operational characteristics, the steps for predicting production capacity and allocating production strategies using deep learning algorithms are as follows:

[0108] The capacity forecasting module uses a deep feedforward neural network (DNN) for modeling, and the calculation process is as follows:

[0109] ,

[0110] ,

[0111] in, For the first The output features of the layer The overall working characteristics of the input layer, For the first Layer weight matrix and bias, This is the final capacity forecast. The total number of layers in the network;

[0112] Using the mean squared error (MSE) as the loss function, it is defined as follows: :

[0113] ,

[0114] The model parameters are then updated using gradient descent, with the following update formula:

[0115] ,

[0116] in, The loss value represents the mean squared error between the predicted and actual values. For the sample size, For the first The predicted value for each sample, For the first The actual production capacity of each sample Network parameters, including weights and biases. For learning rate, The number of iterations is used here. A deep learning model is employed to capture data patterns, thereby improving the accuracy of capacity forecasting.

[0117] In step S4, the allocation module adjusts production capacity according to the production strategy, including the tool type and quantity at each workstation, as well as the tightening sequence and tightening path; at the same time, the error warning module constructs an anomaly detection model based on a variational autoencoder (VAE), using the fused work features as input to analyze abnormal data in the tightening process, including torque anomalies and speed anomalies; and uses unsupervised learning to establish a probability distribution model of normal operation data. During operation, the reconstruction error and distribution deviation of real-time data are used as anomaly indicators, and multi-level warnings are issued in conjunction with business rules.

[0118] The steps for the allocation module to adjust production capacity according to the production strategy are as follows:

[0119] The allocation module adjusts production capacity, including:

[0120] According to the task priority allocation tool, the allocation formula is as follows:

[0121] ,

[0122] in, To be assigned to workstations Tool types, For tools At workstation The fit score on the screen

[0123] The path optimization formula is as follows:

[0124] ,

[0125] in, The total cost of the path. For path distance, For path Task weights For path The allocation module improves capacity utilization through intelligent allocation and path optimization, while the error warning module ensures production stability through anomaly detection.

[0126] The steps for constructing an anomaly detection model based on a variational autoencoder (VAE), using fused working features as input, and analyzing abnormal data during the tightening process are as follows:

[0127] The latent distribution is constructed using a variational autoencoder (VAE), and the construction formula is as follows:

[0128] ,

[0129] in, Representation of the potential space, The mean and standard deviation of the potential distribution are given. It is random noise. Features for reconstruction;

[0130] Specifically, by using multimodal data fusion based on attention mechanisms, key features can be effectively extracted even in the presence of noise or incomplete data, reducing the impact of single-modal data anomalies on prediction and allocation. Meanwhile, the error warning module builds an anomaly detection model based on variational autoencoder (VAE), accurately capturing anomalies in the tightening process by reconstructing error and distribution deviation as dual indicators. It also establishes a probability distribution model of normal operating data through unsupervised learning, flexibly responding to complex anomalies and combining business rules for multi-level warnings, significantly reducing false alarms and missed alarms.

[0131] The steps for using real-time data reconstruction errors and distribution deviations as anomaly indicators, and combining them with business rules for multi-level early warning, are as follows:

[0132] The reconstruction error is calculated using the following formula:

[0133] ,

[0134] in, Reconstruction error, representing the input features. Reconstructed output with autoencoder The differences between them To synthesize the overall working characteristics, the input feature matrix is ​​fused by the attention mechanism. The synthesized features reconstructed by the autoencoder are then processed by the decoder. From potential representation The generated feature matrix, The square of the L2 norm;

[0135] The distribution bias is calculated using the following formula:

[0136] ,

[0137] in, Let be the distribution bias, representing the approximate posterior distribution of the latent space. With prior distribution Kullback-Leibler divergence between them It is an approximate posterior distribution, based on input features. The generated latent distribution is a Gaussian distribution: , The mean vector is an approximate posterior distribution, generated by the encoder. The standard deviation vector of the approximate posterior distribution is generated by the encoder. The prior distribution is , and the standard normal distribution is . , Let be the mean vector of the prior distribution. Let KL be the identity matrix, representing the independent and identically distributed standard normal distribution. The Kullback-Leibler divergence, used to measure the relative entropy difference between two probability distributions, is defined as:

[0138] ,

[0139] By calculating the reconstruction error and distribution deviation The system evaluates whether the input features conform to the normal operating state of the model. If the deviation is too large, it indicates an anomaly. Anomaly indicators together constitute the basis for anomaly detection in the autoencoder.

[0140] Specifically, the anomaly detection and multi-level early warning mechanism based on variational autoencoders can proactively identify potential problems before anomalies occur, thereby preventing possible production accidents.

[0141] In step S5, the feedback module feeds back the abnormal data to the capacity prediction module and updates the deep learning model in step S3.

[0142] The feedback module compares abnormal data with the prediction results and generates optimized samples. These optimized samples are used to update the deep learning model. The feedback module forms a closed-loop system, which enhances the model's ability to handle complex scenarios.

[0143] 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based intelligent tightening tool capacity allocation system, comprising a data acquisition module, a data upload module, a capacity prediction module, an allocation module, an error warning module, a feedback module, and a user interface, characterized in that, The data acquisition module has a built-in controller, which uses PLC and PC-based controllers to collect the working data of the tightening tool, including sensor data, production history and worker operation records; Sensor data includes torque, speed, and temperature and humidity; Production history includes task sequence and tool usage time; Worker operation records include operation frequency and real-time intervention data; The data upload module is used to upload working data to the capacity forecasting module; The capacity prediction module uses deep learning algorithms to predict capacity based on working data and adjusts production strategies accordingly. The allocation module adjusts production capacity according to the production strategy, including the type and quantity of tools at each workstation, as well as the tightening sequence and tightening path. The error warning module is used to analyze abnormal data during the tightening process, including abnormal torque and abnormal speed, and to provide error warnings. The feedback module is used to feed back abnormal data to the capacity prediction module in real time; The user interface is used by operators to monitor and adjust the status of tightening tools; In the data acquisition module, a multimodal fusion model based on an attention mechanism is adopted to dynamically adjust the weight of the work data and generate comprehensive work features. The fused work features are then input into the allocation module and the error warning module. In the error warning module, an anomaly detection model based on variational autoencoder (VAE) is constructed, with the fused working features as input; Unsupervised learning is used to establish a probability distribution model for normal operation data. During operation, the reconstruction error and distribution deviation of real-time data are used as anomaly indicators, and multi-level early warnings are carried out in combination with business rules.

2. The intelligent tightening tool capacity allocation method based on deep learning, based on the intelligent tightening tool capacity allocation system based on deep learning of claim 1, characterized in that, include: Step S1: Collect working data during the tightening process using the data acquisition module. Step S2 involves using a multimodal fusion model based on an attention mechanism to dynamically adjust the weights of the collected work data, generating comprehensive work features. These fused work features are then input into the allocation module and the error warning module. Step S3: The capacity prediction module uses a deep learning algorithm to predict capacity based on the fused working characteristics and adjusts production strategies accordingly. In step S4, the allocation module adjusts production capacity according to the production strategy, including the tool type and quantity at each workstation, as well as the tightening sequence and tightening path; at the same time, the error warning module constructs an anomaly detection model based on a variational autoencoder (VAE), using the fused work features as input to analyze abnormal data in the tightening process, including torque anomalies and speed anomalies; and uses unsupervised learning to establish a probability distribution model of normal operation data. During operation, the reconstruction error and distribution deviation of real-time data are used as anomaly indicators, and multi-level warnings are issued in conjunction with business rules. In step S5, the feedback module sends the abnormal data back to the capacity prediction module to update the deep learning model in step S3.

3. The deep learning-based intelligent tightening tool capacity deployment method of claim 2, wherein: The steps for generating comprehensive work features by using an attention-based multimodal fusion model to dynamically adjust the weights of the collected work data are as follows: Data preprocessing is performed, and after standardization of the multimodal data, preliminary features are extracted using the following formula: ， wherein, is a feature matrix for sensor data, is a feature matrix for production history records, is a feature matrix for worker operation records, is raw data input for a corresponding modality, are a convolutional neural network and a recurrent neural network, respectively; The weights of each modal feature are dynamically adjusted based on an attention mechanism, and the calculation formula is as follows: ， wherein, is a weight of the modality, , , is a preliminary feature of the modality, , is a learnable parameter matrix for weight calculation, is a modality index normalized by weight, taking a value range of ; After fusing modal features according to weights, the final comprehensive feature is generated. The formula for the comprehensive feature is: ， wherein, is a comprehensive working feature, is a weight matrix of the fully connected layer, is a bias of the fully connected layer, is a nonlinear activation function.

4. The deep learning-based intelligent tightening tool capacity deployment method of claim 3, wherein: Based on the fused working characteristics, the steps for predicting production capacity and allocating production strategies using deep learning algorithms are as follows: The capacity prediction module uses a deep feedforward neural network (DNN) for modeling, and the calculation process is as follows: ， ， wherein, is the first output feature of the layer, is the integrated working feature of the input layer, is the first weight matrix and bias of the layer, is the final capacity prediction value, is the total number of layers of the network; Using mean square error (MSE) as the loss function, defined as : ， The model parameters are then updated using gradient descent, with the following update formula: ， wherein, is the loss function, representing the mean squared error between predicted and true values, is the number of samples, is the predicted value for the th sample, is the true capacity value for the th sample, are the network parameters, including weights and biases, is the learning rate, is the number of iterations.

5. The deep learning-based intelligent tightening tool capacity deployment method of claim 4, wherein: The steps for the allocation module to adjust production capacity according to the production strategy are as follows: The allocation module adjusts production capacity, including: According to the task priority allocation tool, the allocation formula is as follows: ， in, To be assigned to workstations Tool types, For tools At workstation The fit score on the screen The path optimization formula is as follows: ， wherein, is the total cost of the path, is the distance of the path , is the task weight of the path , is the execution time of the path .

6. The intelligent tightening tool production capacity allocation method based on deep learning as described in claim 5, characterized in that: The steps for constructing an anomaly detection model based on a variational autoencoder (VAE), and analyzing abnormal data during the tightening process using fused working features as input, are as follows: The latent distribution is constructed using a variational autoencoder (VAE), and the construction formula is as follows: ， in, Representation of the potential space, The mean and standard deviation of the potential distribution are given. It is random noise. Features for reconstruction.

7. The intelligent tightening tool production capacity allocation method based on deep learning as described in claim 6, characterized in that: The steps for using real-time data reconstruction errors and distribution deviations as anomaly indicators, and combining them with business rules for multi-level early warning, are as follows: The reconstruction error is calculated using the following formula: ， in, Reconstruction error, representing the input features. Reconstructed output with autoencoder The differences between them To synthesize the overall working characteristics, the input feature matrix is ​​fused by the attention mechanism. The synthesized features reconstructed by the autoencoder are then processed by the decoder. From potential representation The generated feature matrix, The square of the L2 norm; The distribution bias is calculated using the following formula: ， in, Let be the distribution bias, representing the approximate posterior distribution of the latent space. With prior distribution Kullback-Leibler divergence between them It is an approximate posterior distribution, based on input features. The generated latent distribution is a Gaussian distribution: , The mean vector is an approximate posterior distribution, generated by the encoder. The standard deviation vector of the approximate posterior distribution is generated by the encoder. The prior distribution is , and the standard normal distribution is . , Let be the mean vector of the prior distribution. Let KL be the identity matrix, representing the independent and identically distributed standard normal distribution. The Kullback-Leibler divergence, used to measure the relative entropy difference between two probability distributions, is defined as: ， By calculating the reconstruction error and distribution deviation The system evaluates whether the input features conform to the normal operating state of the model. If the deviation is too large, it indicates an anomaly. Anomaly indicators together constitute the basis for anomaly detection in the autoencoder.

8. The intelligent tightening tool production capacity allocation method based on deep learning as described in claim 7, characterized in that: The feedback module compares the abnormal data with the prediction results and generates optimized samples, which are then used to update the deep learning model.

Images