Tightening quality evaluation method based on deep learning

Abstract

The invention discloses a tightening quality evaluation method based on deep learning, and relates to the technical field of tightening quality evaluation.The method comprises the steps that torque, pre-tightening force, a lubricating state and environmental data in the tightening process are collected in real time, nonlinear fluctuation of friction factors in the repeated tightening process is captured through an LSTM dynamic model, and the torque and pre-tightening force conversion relation is corrected in time; an adaptive learning rate and full connection layer mapping are adopted, so that the model can automatically update parameters under different environmental conditions, and the tightening speed, the preset torque and the lubrication strategy are dynamically adjusted; the environment integration module quantifies temperature, humidity and salt concentration changes, process parameters are corrected in real time, and instantaneous abnormal fluctuation is effectively restrained.

Description

Technical Field

[0001] The present invention relates to the technical field of tightening quality assessment, and particularly to a tightening quality assessment method based on deep learning. Background Art

[0002] In the current bolt and nut tightening quality assessment system, it mainly relies on the calibration of the fixed relationship between the input torque and the pre-tightening force. In actual production, during repeated tightening operations, as the number of tightenings increases, the contact surface morphology of the bolt and nut will change, and the lubricating layer will gradually wear out, resulting in irregular fluctuations in the friction coefficient. There will be a large deviation between the actual pre-tightening force and the theoretical value.

[0003] At the same time, different environmental temperatures and humidities will also affect the tightening process. For example, in the connection operation of offshore wind turbines, drastic changes in environmental conditions will have unpredictable effects on the surface state of materials and the lubrication effect, and the transmission relationship between torque and pre-tightening force is more complex. In response to this situation, most traditional methods adopt standardized lubrication schemes and preset process parameters. Obviously, this will cause the pre-tightening force of some workpieces to be too low or too high in a specific environment; Therefore, there is an urgent need for a tightening quality assessment method based on deep learning to solve such problems. Summary of the Invention

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

[0005] The present invention provides a tightening quality assessment method based on deep learning to solve the problems that the traditional tightening quality assessment scheme cannot effectively evaluate the non-linear changes during repeated tightening and there are deficiencies in the environmental adaptability assessment.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] An embodiment of the present invention provides a tightening quality assessment method based on deep learning, which includes:

[0008] Step S1, adding environmental simulation, and collecting real-time data during the tightening process by using a torque sensor, a pre-tightening force sensor, and temperature, humidity, and salt spray environment parameter sensors;

[0009] Step S2, fusing and preprocessing the real-time data, and extracting quality features; the quality features include changes in the contact surface state and fluctuations in the friction coefficient caused by repeated tightening;

[0010] Step S3, constructing a dynamic model based on the quality features extracted in Step S2. The dynamic model uses an LSTM neural network to fit and predict the non-linear friction coefficient changes caused by surface state changes and environmental impacts during the repeated tightening process, and corrects the conversion relationship between torque and pre-tightening force.

[0011] Step S4: Use the output of the dynamic model in Step S3 to adjust the tightening process parameters in real time, including adjusting the tightening speed, preset torque, and lubricant replenishment strategy;

[0012] Step S5: Build a digital twin platform, integrate the collected real-time data with the model prediction results through simulation, visually monitor the full-cycle status of the tightening process, and generate a real-time evaluation report.

[0013] As a preferred solution of the tightening quality evaluation method based on deep learning according to the present invention, in Step S1, the collected data includes tightening torque, pre-tightening force, contact surface lubrication state, and environmental indicators.

[0014] As a preferred solution of the tightening quality evaluation method based on deep learning according to the present invention, the preprocessing includes data denoising, normalization, and eigenvalue statistics;

[0015] The data preprocessing step further includes statistically analyzing the data of each tightening cycle in the repeated tightening process to quantify the influence of the contact surface morphology and lubricating layer change on the friction coefficient.

[0016] As a preferred solution of the tightening quality evaluation method based on deep learning according to the present invention, the training of the dynamic model is based on historical tightening test data, and the dynamic model includes a model adaptive parameter adjustment unit for automatically updating the model weights under different environmental conditions.

[0017] As a preferred solution of the tightening quality evaluation method based on deep learning according to the present invention, in Step S3, the steps of fitting and predicting are as follows:

[0018] The dynamic model processes the time series data using the following LSTM unit calculation formula:

[0019] i t =σ(W i x t +U i h t-1 +b i ),

[0020] f t =σ(Wfx t +Ufh t-1 +bf),

[0021]

[0022] o t =σ(W o x t +Uo h t-1 +b o ),

[0023] h t =o t ⊙tanh(c t ),

[0024] where, i t represents the input gate output vector, σ(·) is the sigmoid activation function, W i is the input weight matrix of the input gate, x t is the input feature vector at time t, U i is the hidden state weight matrix of the input gate, h t-1 is the hidden state vector of the previous time step, b i is the bias vector of the input gate, f t represents the forget gate output vector, W f is the input weight matrix of the forget gate, U f is the hidden state weight matrix of the forget gate, b f is the bias vector of the forget gate, is the candidate memory vector, tanh(·) is the hyperbolic tangent activation function, W c is the input weight matrix of the candidate memory, U c is the hidden state weight matrix of the candidate memory, b c is the bias vector of the candidate memory, c t is the updated memory state vector, c t-1 is the memory state vector of the previous time step, ⊙ represents the element-wise multiplication operation, o t represents the output gate output vector, W o is the input weight matrix of the output gate, U o is the hidden state weight matrix of the output gate, b o is the bias vector of the output gate, h t is the hidden state vector at the current time step;

[0025] The network output is mapped to the friction factor correction amount through a fully connected layer, and the formula is:

[0026] y t =W y h t +b y ,

[0027] where, y t is the friction factor correction amount of the network output, W y is the mapping matrix of the fully connected layer, h t is the hidden state vector at the current time step, b y is the bias vector of the fully connected layer.

[0028] As a preferred solution of the method for evaluating tightening quality based on deep learning according to the present invention, wherein: in step S3, the step of fitting and predicting further includes:

[0029] Using the prediction result to correct the torque, and the relationship is expressed as:

[0030] τ′ t =τ t +λy t ,

[0031] wherein, τ′ t is the corrected torque, τ t is the original torque, λ is the adjustment coefficient, and y t is the friction factor correction amount;

[0032] Using the mean square error loss function for model training, and the loss function formula is:

[0033]

[0034] wherein, is the mean square error loss, N is the number of samples, y t is the friction factor correction amount predicted by the network, is the actual friction factor change value;

[0035] Introducing an adaptive learning rate, and the adaptive learning rate calculation formula is:

[0036] α t =α0exp(-γΔE t ),

[0037] wherein, α t is the adaptive learning rate at the current moment, α0 is the initial learning rate, exp(·) represents the exponential function, γ is the decay factor, and ΔE t is the current error change rate, which is calculated based on the error difference between the previous and current moments.

[0038] As a preferred solution of the method for evaluating tightening quality based on deep learning according to the present invention, wherein: in step S5, the method of integrating the collected real-time data with the model prediction result is: integrating the real-time data collected by the sensor with the model output, and defining the simulation state vector as s t :

[0039] s t =αfz t +βfy t ,

[0040] wherein, S trepresents the simulation state vector at time t, α f is the sensor data fusion weight, z t is the sensor data vector collected at time t, β f is the model prediction data fusion weight, y t The predicted value output by the model;

[0041] The deviation between real-time data and model prediction is calculated to quantify the state difference of the tightening process. The calculation formula is: d t =|z t -y t |, where d t is the deviation at time t, |·| represents the norm operation;

[0042] Generate an alarm signal based on the deviation: If d t >δ, then A t =1, if d t ≤δ, then A t =0, where A t It is an alarm flag, a value of 1 indicates that the deviation exceeds the preset threshold, and δ is the deviation threshold parameter.

[0043] As a preferred solution of the tightening quality assessment method based on deep learning described in the present invention, in step S5, the digital twin platform calibrates the data according to a unified update cycle, and the update cycle is recorded as T s , expressed as:

[0044] z t =z(tT s ), y t =y(tT s ),

[0045] Among them, T s Indicates the data update cycle;

[0046] The platform calculates and reports indicators using the following formula:

[0047] R t =ω|s t |+ηA t ,

[0048] Among them, R t represents the reporting index at time t, ω is the weight factor of the simulation state norm, and η is the weight factor of the alarm flag.

[0049] As a preferred embodiment of the method for evaluating tightening quality based on deep learning according to the present invention, wherein: the digital twin platform is constructed by using virtual simulation technology, and digitally replicates the tightening process based on the feedback of real-time data, so as to timely remind on-site operators when there are deviations in the pre-tightening force and friction state;

[0050] The environmental parameter sensor operates synchronously with the tightening device by using an integrated module, and dynamically adjusts the lubrication and tightening process parameters under conditions of drastic changes in temperature and humidity or salt spray erosion.

[0051] As a preferred embodiment of the method for evaluating tightening quality based on deep learning according to the present invention, wherein: in the digital twin platform, the integrated module of the environmental parameter sensor dynamically adjusts the lubrication and tightening process parameters by collecting environmental data in real time and combining with an adjustment algorithm under conditions of drastic changes in temperature and humidity or salt spray erosion. Specifically:

[0052] The integrated module collects the temperature T t , humidity H t and salt concentration S t environmental data, and calculates the environmental severity factor E t , and the calculation formula is:

[0053]

[0054] wherein, T t represents the current temperature, H t represents the current humidity, S t represents the current salt concentration, α T is the temperature sensitivity weight, α H is the humidity sensitivity weight, α s is the salt concentration sensitivity weight, T ref , H ref and S ref are the reference environmental temperature, humidity and salt concentration respectively, and exp(·) is the exponential function;

[0055] Based on the environmental factor, dynamically adjust the lubrication parameter, and the adjustment formula is:

[0056] L t = L0(1 + k L E t ),

[0057] wherein, L t is the adjusted lubrication parameter, L0 is the reference lubrication parameter, and k L is the lubrication parameter adjustment coefficient,

[0058] Correct the tightening process parameter, and the formula is:

[0059] P t= P0(1 - k P E t ),

[0060] where P t is the corrected tightening process parameter, P0 is the reference tightening process parameter, and k P is the tightening parameter adjustment coefficient;

[0061] Using moving average filtering, the environmental severity factor is smoothed in the time domain, and the expression is:

[0062]

[0063] where is the smoothed environmental factor, n is the time window length, and i represents the index of the time series.

[0064] The beneficial effects of the present invention are as follows: The present invention collects torque, pre-tightening force, lubrication state, and environmental data in real time during the tightening process, uses the LSTM dynamic model to capture the non-linear fluctuations of the friction coefficient in the repeated tightening process, and timely corrects the torque-pre-tightening force conversion relationship; adopts an adaptive learning rate and a fully connected layer mapping to enable the model to automatically update parameters under different environmental conditions, dynamically adjust the tightening speed, preset torque, and lubrication strategy; the environmental integration module quantifies the changes in temperature, humidity, and salt concentration, and corrects the process parameters in real time to effectively suppress instantaneous abnormal fluctuations.

[0065] The present invention effectively alleviates the problems of insufficient non-linearity and environmental adaptability of traditional methods in repeated tightening, and realizes high-precision, real-time, and intelligent tightening quality evaluation and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0067] Figure 1 It is a schematic flow chart of the tightening quality evaluation method based on deep learning in Embodiment 1. DETAILED DESCRIPTION OF THE INVENTION

[0068] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the specific embodiments of the present invention in detail with reference to the drawings of the specification.

[0069] In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways than those specifically described herein. Those skilled in the art can make similar generalizations without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0070] Secondly, as used herein, "an embodiment" or "embodiments" refer to specific features, structures, or characteristics that may be included in at least one implementation of the present invention. The phrase "in an embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it an individual or alternative embodiment that is mutually exclusive of other embodiments.

[0071] Embodiment 1, referring to Figure 1 , this embodiment provides a method for evaluating tightening quality based on deep learning, including:

[0072] Step S1, adding environmental simulation, and using a torque sensor, a pre-tightening force sensor, and temperature, humidity, and salt spray environment parameter sensors to collect real-time data during the tightening process;

[0073] The collected data includes tightening torque, pre-tightening force, contact surface lubrication state, and environmental indicators;

[0074] Step S2, fusing and preprocessing the real-time data to extract quality features; the quality features include changes in the contact surface state and fluctuations in the friction coefficient caused by repeated tightening;

[0075] The preprocessing includes data denoising, normalization, and eigenvalue statistics;

[0076] The data preprocessing step further includes statistically analyzing the data of each tightening cycle during the repeated tightening process to quantify the influence of changes in the contact surface morphology and lubricating layer on the friction coefficient;

[0077] Step S3, constructing a dynamic model based on the quality features extracted in step S2. The dynamic model uses an LSTM neural network to fit and predict the non-linear changes in the friction coefficient caused by surface state changes and environmental influences during the repeated tightening process, and corrects the conversion relationship between torque and pre-tightening force;

[0078] The training of the dynamic model is based on historical tightening test data. The dynamic model includes a model adaptive parameter adjustment unit for automatically updating the model weights under different environmental conditions;

[0079] In step S3, the steps of fitting and predicting are as follows:

[0080] The dynamic model processes time series data using the following LSTM unit calculation formula:

[0081] i t =σ(Wi x t +U i h t-1 +b i ),

[0082] g t =σ(W f x t +U f h t-1 +b f ),

[0083]

[0084] o t =σ(W o x t +U o h t-1 +b o ),

[0085] h t =o t ⊙tanh(c t ),

[0086] where i t represents the input gate output vector, σ(·) is the sigmoid activation function, W i is the input weight matrix of the input gate, x t is the input feature vector at time t, U i is the hidden state weight matrix of the input gate, h t-1 is the hidden state vector of the previous time step, b i is the bias vector of the input gate, f t represents the forget gate output vector, W f is the input weight matrix of the forget gate, U f is the hidden state weight matrix of the forget gate, b f is the bias vector of the forget gate, is the candidate memory vector, tanh(·) is the hyperbolic tangent activation function, W c is the input weight matrix of the candidate memory, U c is the hidden state weight matrix of the candidate memory, b c is the bias vector of the candidate memory, c t is the updated memory state vector, c t-1 is the memory state vector of the previous time step, ⊙ represents the element-wise multiplication operation, o t represents the output gate output vector, W o is the input weight matrix of the output gate, U o is the hidden state weight matrix of the output gate, b ois the bias vector of the output gate, h t is the hidden state vector at the current moment;

[0087] The network output is mapped to the friction factor correction amount through a fully connected layer, and the formula is:

[0088] y t = W y h t + b y ,

[0089] where y t is the friction factor correction amount output by the network, W y is the fully connected layer mapping matrix, h t is the hidden state vector at the current moment, b y is the fully connected layer bias vector;

[0090] In step S3, the steps of fitting and prediction further include:

[0091] Using the prediction result to correct the torque, and the relationship is expressed as:

[0092] τ′ t = τ t + λy t ,

[0093] where τ′ t is the corrected torque, τ t is the original torque, λ is the adjustment coefficient, and y t is the friction factor correction amount;

[0094] The mean square error loss function is used for model training, and the loss function formula is:

[0095]

[0096] where, is the mean square error loss, N is the number of samples, y t is the friction factor correction amount predicted by the network, is the actual friction factor change value;

[0097] An adaptive learning rate is introduced, and the adaptive learning rate calculation formula is:

[0098] α t = α0exp(-γΔE t ),

[0099] where α t is the adaptive learning rate at the current moment, α0 is the initial learning rate, exp(·) represents the exponential function, γ is the decay factor, and ΔE tis the current error change rate, calculated based on the error difference between the previous and current moments;

[0100] Specifically, an LSTM neural network dynamic model is constructed, and the time-series quality features during the tightening process are input into the network for processing. The model uses the time-series structure composed of an input gate, a forget gate, a candidate memory, and an output gate to capture the non-linear fluctuations of the friction coefficient caused by changes in surface conditions and environmental factors;

[0101] The fully connected layer mapping and subsequent torque correction provide a correction basis for the actual torque data, adjust the conversion relationship in real time, and the mean square error loss function ensures the minimization of the error between the predicted value and the actual situation during the model training process. The adaptive learning rate mechanism dynamically adjusts the weight update rate for different working conditions and can respond to subtle changes during multiple tightening operations, thereby improving the evaluation accuracy of the overall tightening process;

[0102] Step S4: Use the output of the dynamic model in Step S3 to adjust the tightening process parameters in real time, including adjusting the tightening speed, preset torque, and the strategy of supplementing lubricant;

[0103] Step S5: Construct a digital twin platform, integrate the collected real-time data with the model prediction results through simulation, visually monitor the full-cycle state of the tightening process, and generate a real-time evaluation report;

[0104] In Step S5, the method of integrating the collected real-time data with the model prediction results is as follows: Integrate the real-time data collected by the sensor with the model output, and define the simulation state vector as s t :

[0105] s t =α f z t +β f y t ,

[0106] where s t represents the simulation state vector at time t, α f is the sensor data fusion weight, z t is the sensor data vector collected at time t, β f is the model prediction data fusion weight, and y t is the predicted value output by the model;

[0107] Calculate the deviation between the real-time data and the model prediction to quantify the state difference of the tightening process. The calculation formula is: d t =|z t -y t |, where d t is the deviation at time t, and |·| represents the norm operation;

[0108] Generate an alarm signal based on the deviation: If d t > δ, then A t = 1. If d t ≤ δ, then A t = 0. Here, A t is the alarm flag. A value of 1 indicates that the deviation exceeds the preset threshold, and δ is the deviation threshold parameter;

[0109] In step S5, the digital twin platform calibrates the data according to a unified update period, and the update period is denoted as T s , expressed as:

[0110] z t = z(t - T s ), y t = y(t - T s ),

[0111] where T s represents the data update period;

[0112] The platform calculates the report index, and the calculation formula is:

[0113] R t = ω|s t | + ηA t ,

[0114] where R t represents the report index at time t, ω is the weight factor of the simulation state norm, and η is the weight factor of the alarm flag;

[0115] Specifically, the digital twin platform organically integrates real-time data and model prediction results, constructs a simulation state using data fusion, realizes dynamic replication of the full-cycle process. The deviation calculation and alarm mechanism enable the system to promptly reflect the deviation of the pre-tightening force and friction state. Data synchronization ensures the consistency of data for each module. The report index synthesizes the simulation state and abnormal alerts, providing intuitive monitoring information and evaluation basis for on-site operators;

[0116] The digital twin platform is constructed using virtual simulation technology, digitally replicates the tightening process based on the feedback of real-time data, and thus promptly alerts on-site operators when there are deviations in the pre-tightening force and friction state;

[0117] The environmental parameter sensor operates synchronously with the tightening device using an integrated module, and dynamically adjusts lubrication and tightening process parameters under conditions of drastic temperature and humidity changes or salt spray erosion;

[0118] In the digital twin platform, the environmental parameter sensor integrated module dynamically adjusts lubrication and tightening process parameters under conditions of drastic temperature and humidity changes or salt spray erosion by collecting environmental data in real time and combining adjustment algorithms. Specifically:

[0119] The integrated module collects the temperature T t , humidity H t and salt concentration S t of environmental data and calculates the environmental severity factor E t . The calculation formula is:

[0120]

[0121] where T t represents the current temperature, H t represents the current humidity, S t represents the current salt concentration, α T is the temperature sensitivity weight, α H is the humidity sensitivity weight, α s is the salt concentration sensitivity weight, T ref , H ref and S ref are the reference environmental temperature, humidity and salt concentration respectively, and exp(·) is the exponential function;

[0122] Based on the environmental factor, the lubrication parameters are dynamically adjusted. The adjustment formula is:

[0123] L t = L0(1 + k L E t ),

[0124] where L t is the adjusted lubrication parameter, L0 is the reference lubrication parameter, and k L is the lubrication parameter adjustment coefficient;

[0125] The tightening process parameters are corrected. The formula is:

[0126] P t = P0(1 - K P E t ),

[0127] where P t is the corrected tightening process parameter, P0 is the reference tightening process parameter, and k P is the tightening parameter adjustment coefficient;

[0128] The sliding average filter is used to smooth the environmental severity factor in the time domain. The expression is:

[0129]

[0130] where is the smoothed environmental factor, n is the length of the time window, and i represents the index of the time series;

[0131] Specifically, the integration module quantifies the environmental impact under the drastic changes of temperature, humidity and salt concentration by collecting environmental parameters in real time and constructing an environmental severity factor using an exponential function. Based on this factor, the lubrication and tightening process parameters are dynamically adjusted using a linear correction method. The moving average filter can suppress instantaneous abnormal fluctuations, automatically respond to environmental changes under special working conditions, effectively improve the stability and safety of lubrication and tightening operations, and at the same time provide intuitive data support for on-site operators, enhancing the adaptive ability and real-time monitoring effect of the overall process.

[0132] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A tightening quality evaluation method based on deep learning, characterized in that: Including, Step S1: Add environmental simulation, and collect real-time data during the tightening process using a torque sensor, a pre-tightening force sensor, and sensors for temperature, humidity, and salt spray environment parameters. Step S2: Fuse and preprocess the real-time data to extract quality features; the quality features include changes in the contact surface state and fluctuations in the friction coefficient caused by repeated tightening. Step S3: Build a dynamic model based on the quality features extracted in Step S2. The dynamic model uses an LSTM neural network to fit and predict the non-linear change in the friction coefficient caused by surface state changes and environmental impacts during the repeated tightening process, and corrects the conversion relationship between torque and pre-tightening force. Step S4: Use the output of the dynamic model in Step S3 to adjust the tightening process parameters in real time, including strategies for adjusting the tightening speed, preset torque, and replenishing lubricant. Step S5: Build a digital twin platform, integrate the collected real-time data with the model prediction results through simulation, visually monitor the full-cycle state of the tightening process, and generate a real-time evaluation report.

2. The method for evaluating tightening quality based on deep learning according to claim 1, characterized in that: In Step S1, the collected data includes tightening torque, pre-tightening force, contact surface lubrication state, and environmental indicators.

3. The method for evaluating tightening quality based on deep learning according to claim 1, characterized in that: The preprocessing includes data denoising, normalization, and eigenvalue statistics. The preprocessing further includes statistical analysis of the data for each tightening cycle during the repeated tightening process to quantify the impact of changes in the contact surface morphology and lubricating layer on the friction coefficient.

4. The method for evaluating tightening quality based on deep learning according to claim 1, characterized in that: The training of the dynamic model is based on historical tightening test data. The dynamic model includes a model adaptive parameter adjustment unit for automatically updating the model weights under different environmental conditions.

5. The method for evaluating tightening quality based on deep learning according to claim 4, wherein: In Step S3, the steps for fitting and predicting are as follows: The dynamic model uses the following LSTM unit calculation formula to process time series data: i t = σ(W i x t + U i h t-1 + b i ), f t = σ(W f x t + U f h t-1 + b f ), o t = σ(W o x t + U o h t-1 + b o ), h t = o t ⊙tanh(c t ) where, i t represents the input gate output vector, σ(·) is the sigmoid activation function, W i is the input weight matrix of the input gate, x t is the input feature vector at time t, U i is the hidden state weight matrix of the input gate, h t-1 is the hidden state vector of the previous time step, b i is the bias vector of the input gate, f t represents the forget gate output vector, W f is the input weight matrix of the forget gate, U f is the hidden state weight matrix of the forget gate, b f is the bias vector of the forget gate, is the candidate memory vector, tanh(·) is the hyperbolic tangent activation function, W c is the input weight matrix of the candidate memory, U c is the hidden state weight matrix of the candidate memory, b c is the bias vector of the candidate memory, c t is the updated memory state vector, c t-1 is the memory state vector of the previous time step, ⊙ represents the element-wise multiplication operation, o t represents the output gate output vector, W o is the input weight matrix of the output gate, U o is the hidden state weight matrix of the output gate, b o is the bias vector of the output gate, h t is the hidden state vector at the current time step; The network output is mapped to the friction coefficient correction amount through a fully connected layer, and the formula is: y t = W y h t + b y , Among them, y t is the friction factor correction amount output by the network, W y is the fully connected layer mapping matrix, h t is the hidden state vector at the current moment, b y is the fully connected layer bias vector.

6. The method for evaluating tightening quality based on deep learning according to claim 5, wherein: In Step S3, the steps for fitting and predicting also include: Use the prediction result to correct the torque, and the relationship expression is: τ′ t = τ t + λy t , Among them, τ′ t is the corrected torque, τ t is the original torque, λ is the adjustment coefficient, and y t is the friction factor correction amount; Use the mean square error loss function for model training, and the loss function formula is: Among them, is the mean square error loss, N is the number of samples, and y t is the correction amount of the friction factor predicted by the network, is the actual change value of the friction factor; Introduce an adaptive learning rate, and the adaptive learning rate calculation formula is: α t = α0exp(-γΔE t ) where α t is the adaptive learning rate at the current moment, α0 is the initial learning rate, exp(·) represents the exponential function, γ is the decay factor, and ΔE t is the current error change rate, which is calculated based on the error difference between the previous and current moments.

7. The method for evaluating tightening quality based on deep learning according to claim 1, characterized in that: In step S5, the method of integrating the collected real-time data with the model prediction results is as follows: integrate the real-time data collected by the sensor with the model output, and define the simulation state vector as s t : s t = α f z t + β f y t , Among them, s t represents the simulation state vector at time t, α f is the weight of sensor data fusion, z t is the sensor acquisition data vector at time t, β f is the weight of model prediction data fusion, y t is the predicted value output by the model; Quantify the state difference of the tightening process by calculating the deviation between the real-time data and the model prediction. The calculation formula is: d t = |z t - y t |, where d t is the deviation at time t, and |·| represents the norm operation; Generate an alarm signal based on the deviation amount: If d t > δ, then A t = 1. If d t ≤ δ, then A t = 0. Here, A t is the alarm flag. A value of 1 indicates that the deviation exceeds the preset threshold, and δ is the deviation threshold parameter.

8. The method for evaluating tightening quality based on deep learning according to claim 7, wherein: In step S5, the digital twin platform calibrates the data according to a unified update period, and the update period is denoted as T s , which is expressed as: z t = z(t - T s ), y t = y(t - T s ), Among them, T s represents the data update period; The platform calculates the report index, and the calculation formula is: R t = ω|s t | + ηA t , where R t represents the reporting index at time t, ω is the weight factor of the simulation state norm, and η is the weight factor of the alarm flag.

9. The method for evaluating tightening quality based on deep learning according to claim 8, wherein: The digital twin platform is built using virtual simulation technology, and digitally replicates the tightening process based on the feedback of real-time data. The environmental parameter sensor uses an integrated module to run synchronously with the tightening device, and dynamically adjusts the lubrication and tightening process parameters under conditions of sudden changes in temperature and humidity or salt spray erosion.

10. The method for evaluating tightening quality based on deep learning according to claim 9, wherein: In the digital twin platform, under conditions of sudden changes in temperature and humidity or salt spray erosion, the environmental parameter sensor integrated module dynamically adjusts the lubrication and tightening process parameters by collecting real-time environmental data and combining with an adjustment algorithm. Specifically: The integrated module collects the temperature T t , humidity H t and salt concentration S t environmental data and calculates the environmental severity factor E t , and the calculation formula is: Among them, T t represents the current temperature, H t represents the current humidity, S t represents the current salt concentration, α T is the temperature sensitivity weight, α H is the humidity sensitivity weight, α S is the salt concentration sensitivity weight, T ref , H ref and S ref are the reference environmental temperature, humidity and salt concentration respectively, and exp(·) is the exponential function; Dynamically adjust the lubrication parameters based on the environmental factor, and the adjustment formula is: L t = L0(1 + k L E t )) Among them, L t is the adjusted lubrication parameter, L0 is the reference lubrication parameter, and k L is the lubrication parameter adjustment coefficient. Correct the tightening process parameters, and the formula is: P t = P0(1 - k P E t ), Among them, P t is the corrected tightening process parameter, P0 is the reference tightening process parameter, and k P is the tightening parameter adjustment coefficient; Use a moving average filter to perform time-domain smoothing on the environmental severity factor, and the expression is: Among them, is the smoothed environmental factor, n is the length of the time window, and i represents the index of the time series.

Images