Multi-mode intelligent tension self-adaptive control method

Abstract

The invention discloses a multi-mode intelligent tension self-adaptive control method, and belongs to the technical field of tension control of wire rod processing devices. According to the method, an LSTM-Transform hybrid model is utilized to realize the collaborative optimization control of tension prediction, the rotating speed and the pitch; by means of an edge computing architecture and a Modbus-TCP real-time communication protocol, a data processing and control instruction generation period is compressed to be within 50ms, and the dynamic response capability of a system is improved; and meanwhile, a graded safety protection mechanism is designed, a robust control mode is switched to at a millisecond level when vibration energy exceeds a limit or the temperature is abnormal, and the MES system is linked to push a fault code, so that the control precision is remarkably improved, the material loss rate and the operation and maintenance complexity are reduced, and a high-reliability and self-adaptive intelligent solution is provided for high-speed wire rod processing equipment.

Description

Technical Field

[0001] The present invention relates to a tension control method, in particular to a multi-modal intelligent tension adaptive control method, belonging to the technical field of tension control of wire processing devices. Background Art

[0002] Real-time tension control in wire processing equipment is a core process in wire manufacturing. Its control accuracy and stability directly impact the mechanical properties and surface quality of stranded wire. Currently, the industry generally uses an indirect control solution based on a magnetic powder brake, inferring tension using a drag torque model. However, in practice, this is susceptible to dynamic factors such as fluctuations in the material's friction coefficient and capstan wear, resulting in control errors as high as ±15%. Direct detection using contact sensors, while improving measurement reliability, can also lead to deviations in sensor mounting angles due to mechanical structural limitations, introducing additional errors due to the pulley's own weight. Furthermore, high-speed spindle vibrations can superimpose high-frequency noise, resulting in a signal-to-noise ratio below 10dB.

[0003] Existing systems mostly use a decentralized control architecture, resulting in a 50-100ms delay between sensor signal processing, motor speed regulation, and pitch adjustment modules. This requires periodic manual calibration of sensor parameters to compensate for temperature drift, making the process cumbersome and time-consuming. Furthermore, there is no real-time compensation mechanism for the coupling effect of dynamic pulley inclination and mechanical vibration interference. Traditional low-pass filtering methods struggle to effectively isolate valid signals, and safety response delays exceeding 200ms in the event of a sudden disconnection or overload, resulting in material waste exceeding 5%.

[0004] The above technical bottlenecks have put forward clear improvement requirements for high-precision and high-robustness tension control, requiring a new control solution that integrates multi-source data fusion processing, dynamic error compensation and intelligent collaborative decision-making. Summary of the Invention

[0005] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a multi-modal intelligent tension adaptive control method.

[0006] Technical solution: The multi-modal intelligent tension adaptive control method of the present invention comprises the following steps:

[0007] Acquire multi-source data from wire processing devices in real time, including tension sensor data, vibration time-domain signals, pulley inclination data, bearing temperature data, and process parameter data, and perform fast Fourier transform on the vibration time-domain signals to generate vibration spectrum data;

[0008] Preprocessing of multi-source data includes: compensating tension sensor data; using an improved empirical mode decomposition algorithm to separate vibration noise and retain qualified IMF components; and correcting the cantilever beam tension sensor data using a bivariate model.

[0009] Build an LSTM-Transformer dual-output network. The input is preprocessed multi-source data, and the output includes two items: the first output item is the real-time tension value of the processed wire within a certain period of time in the future, and the second output item is the speed adjustment value of the stranding drive motor of the wire processing device and the pitch adjustment value of the stranding pitch adjustment mechanism.

[0010] The output data of the LSTM-Transformer dual-output network is converted into a pulse-width modulated signal to drive the servo motor and synchronized with the programmable logic controller to adjust the pitch hydraulic valve to control the pitch parameters of the wire stranding process.

[0011] Detect whether the current state meets the control conditions of the robust control mode. If yes, switch to the preset robust control mode and trigger the sound and light alarm.

[0012] Furthermore, the formula for compensating the tension sensor data is:

[0013]

[0014] Where, F true Indicates the actual tension value after compensation, F raw represents the original tension value, M represents the weight of the pulley, θ represents the real-time pulley inclination angle, which is used to compensate for the measurement error when the pulley rotates, ω represents the speed of the capstan, r represents the distance from the center of the pulley to the center of the main shaft of the wire processing equipment, and φ represents the angle between the centrifugal force and the wire tension measurement direction.

[0015] Furthermore, the improved empirical mode decomposition algorithm is used to separate the vibration noise, and the steps of retaining the IMF components that meet the conditions include:

[0016] Determine the time delay δt between the vibration signal and the tension signal through cross-correlation analysis, and align the time axes of the vibration signal and the tension signal according to the time delay δt;

[0017] Perform 5-layer empirical mode decomposition on the tension signal F(t) to obtain IMF1 to IMF5 components;

[0018] Calculate the Pearson correlation coefficient r between each IMF component and the vibration signal a(t) i , mark|r i IMF components with |>0.7 are noise-dominated modes, where i=1,2,3,4,5;

[0019] Wavelet soft thresholding is performed on the noise-dominated mode, and the threshold is calculated as Where σ1 is the noise standard deviation, N is the signal length;

[0020] Retain Pearson correlation coefficient | r i|<0.5 IMF component, synthesize the tension signal F after noise reduction vib_comp .

[0021] Furthermore, the structure of the LSTM-Transformer dual-output network includes an input layer, two bidirectional LSTM layers, a multi-head attention layer and a fully connected layer, where the input layer is used to receive a 28-dimensional feature vector, which contains 6-channel time-series tension sensing data, vibration spectrum data, temperature data, inclination data, and speed deviation data.

[0022] Furthermore, the reward function of the LSTM-Transformer dual-output network is:

[0023] R=10 / (1+σ2)-0.1∣Δn∣-0.05∣Δp∣,

[0024] Where σ2 is the standard deviation of tension, Δn is the speed adjustment, and Δp is the pitch adjustment.

[0025] Furthermore, the step of detecting whether the current state satisfies the control conditions of the robust control mode, and if so, switching to the preset robust control mode includes:

[0026] Establish control conditions, including:

[0027] The first condition is that the RMS value of the triaxial vibration acceleration in any direction exceeds 5g and the duration is 200ms;

[0028] The second condition is that the bearing temperature is greater than 85°C or the temperature rise rate is greater than 2°C / min;

[0029] Detect the current state. If the first or second condition is met, switch the action, including:

[0030] The output items of the LSTM-Transformer dual-output network are frozen, and the PID controller is switched to control the servo motor. The speed adjustment range is limited to ±1 rpm, and the pitch adjustment is frozen. A red alarm is triggered on the human-machine interface, and the main operation screen displays a red flashing border. A three-level alarm dialog box pops up, containing the fault code and occurrence time. The fault code is then pushed to the manufacturing execution system, which automatically generates a maintenance work order and assigns a maintenance task.

[0031] Furthermore, the expression of the bivariate model is:

[0032]

[0033] Where △F represents the temperature drift compensation of the tension measurement value, k1 and k2 represent the static temperature drift coefficient and dynamic temperature drift coefficient, respectively, T represents the current temperature of the tension sensor, T0 represents the ambient temperature, and dT / dt represents the temperature change rate.

[0034] Furthermore, the step of correcting the tension sensor data of the cantilever beam using the bivariate model includes:

[0035] Monitor bearing temperature in real time and activate dynamic compensation when the temperature change rate exceeds the preset threshold;

[0036] Correct zero drift according to the calibration curve.

[0037] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0038] 1. This invention optimizes common tension control issues in traditional wire processing devices through the collaborative design of multimodal sensor fusion and intelligent control architecture. Specifically, by collecting multi-source data such as tension, vibration, inclination, and temperature, combined with a dynamic compensation algorithm, it effectively reduces measurement errors caused by mechanical structure deviations and environmental interference, thereby improving the stability and reliability of tension control.

[0039] 2. The present invention uses machine learning methods to optimize control parameters, achieve coordinated adjustment of speed and pitch, maintain rapid response capabilities under complex working conditions, and reduce the risk of wire breakage;

[0040] 3. The present invention simplifies the operation process through efficient communication protocols and modular design, making it convenient for non-professionals to use;

[0041] 4. The present invention monitors key indicators in real time, automatically switches to a robust control mode and triggers an early warning in abnormal situations to ensure production safety;

[0042] 5. In addition, the present invention supports automatic calibration and offline debugging functions, which reduces maintenance requirements, improves equipment utilization, and takes into account practicality and economy while ensuring accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the present invention;

[0044] Figure 2 It is a flowchart of the present invention;

[0045] Figure 3 A flowchart for denoising the IMF components of the noise-dominant mode;

[0046] Figure 4 This is the workflow diagram of the LSTM-Transformer dual-output network. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of this application more clear, this application is further described in detail below with reference to the accompanying drawings and embodiments.

[0048] The multi-modal intelligent tension adaptive control method described in this embodiment is combined with Figures 1 to 2 , the method comprises the following steps:

[0049] Step 1: Acquire multi-source data of the wire processing device in real time, including tension sensor data, vibration time domain signal, pulley inclination data, bearing temperature data, and process parameter data, and perform fast Fourier transform on the vibration time domain signal to generate vibration spectrum data.

[0050] Before acquiring multi-source data, sensors must be pre-deployed, including a 6-channel tension sensor, a triaxial vibration sensor, an inclination sensor, and a PT100 temperature sensor. Tension sensing data refers to the real-time measurement of wire tensile force by a 6-channel tension sensor pre-installed on the wire processing device. Vibration time-domain data is collected by a vibration sensor installed on the wire processing device's main shaft to reflect mechanical vibration interference. Pulley inclination data is obtained from the pulley assembly's inclination sensor. Bearing temperature data refers to the temperature rise of the monitoring device's transmission system. Process parameters, such as sheave speed, lay direction, pitch, and number of wires processed, are sampled at a frequency of ≥1000Hz.

[0051] Step 2: Preprocess the multi-source data, including:

[0052] Compensate tension sensor data;

[0053] An improved empirical mode decomposition algorithm is used to separate vibration noise and retain qualified IMF components;

[0054] And the tension sensing data of the cantilever beam is corrected by a two-variable model.

[0055] During the data collection process, measurement errors may occur due to mechanical structure deviations and environmental interference, so the data needs to be compensated to improve the stability and reliability of tension control.

[0056] Furthermore, the formula for compensating the tension sensor data using the inclination angle is:

[0057]

[0058] Where, F true Indicates the actual tension value after compensation, F raw represents the original tension value, that is, the wire tensile force measured in real time by the tension sensor, which is also the tension sensing data described in step 1. M represents the weight of the pulley. θ represents the real-time pulley inclination angle, which is used to compensate for the measurement error when the pulley rotates. ω represents the speed of the capstan. r represents the distance from the center of the pulley to the center of the main shaft of the wire processing equipment. φ represents the angle between the centrifugal force and the wire tension measurement direction.

[0059] Furthermore, the improved empirical mode decomposition algorithm is used to separate the vibration noise, and the steps of retaining the IMF components that meet the conditions include:

[0060] Determine the time delay δt between the vibration signal and the tension signal through cross-correlation analysis, and align the time axes of the vibration signal and the tension signal according to the time delay δt;

[0061] Perform 5-layer empirical mode decomposition on the tension signal F(t) to obtain IMF1 to IMF5 components;

[0062] Calculate the Pearson correlation coefficient r between each IMF component and the vibration signal a(t) i , mark|r i IMF components with |>0.7 are noise-dominated modes, where i=1,2,3,4,5;

[0063] Perform wavelet soft threshold processing on the noise dominant mode to obtain the denoised IMF component;

[0064] Combine Figure 3 After performing wavelet soft threshold denoising, it is necessary to perform a secondary correlation evaluation on the IMF component of the noise-dominant mode. The specific steps are as follows:

[0065] S21, correlation retest after noise IMF denoising:

[0066] The noise IMF component that has completed the wavelet soft threshold processing, that is, the original correlation coefficient |r i |>0.7 IMF, recalculate its Pearson correlation coefficient r with the vibration signal a(t) i If the recalculated correlation coefficient satisfies |r i |<0.5, then the IMF component is included in the valid IMF component set C valid ; If not satisfied | r i |<0.5, the IMF component is directly discarded and does not participate in the subsequent synthesis;

[0067] S22, the final composition of the effective IMF set:

[0068] Valid IMF component set C valid It contains the following two types of components: the first type is the original low-correlation IMF, which has the Pearson correlation coefficient with the vibration signal after the initial EMD decomposition | r i The unprocessed IMF component with |<0.5 retains the original data form directly; the second type is the IMF that meets the standard after denoising, which is initially judged to be noise-dominated, that is, |r i |>0.7 IMF components, after wavelet soft threshold denoising, retest correlation coefficient |r i The noise energy of the correction component with |<0.5 has been significantly suppressed.

[0069] S23, signal synthesis rules:

[0070] Retain Pearson correlation coefficient | r i |<0.5, in order to screen out the IMF component set C with low correlation with vibration noise valid ={IMF k ∣|r k |<0.5, k∈{1,2,3,4,5}}, extract the final residual component R(t) of EMD decomposition, perform signal-to-noise ratio weighted superposition on the effective IMF component, and obtain the denoised tension signal F vib_comp (t), the formula is:

[0071]

[0072] Where, γ is the residual gain coefficient; w k is the weight coefficient, according to the signal-to-noise ratio SNR of each IMF component k Dynamic setting, the calculation formula is:

[0073]

[0074] Where ε is a small constant to prevent zero.

[0075] The process of performing wavelet soft thresholding on the noise-dominated mode includes:

[0076] The Daubechies 4 wavelet basis is used to analyze the dominant mode of the noise, namely |r i The IMF component set with |>0.7 is decomposed into three layers of discrete wavelet to obtain the detail coefficients d of each layer. j,k ;

[0077] The thresholds of each layer are determined according to Donoho's universal threshold criterion. The threshold calculation formula is:

[0078]

[0079] Where, σ j is the estimated noise standard deviation of the j-th layer wavelet coefficient, which is calculated by dividing the median absolute deviation of the j-th layer detail coefficient by 0.6745; N is the signal length;

[0080] Extract the detail coefficients d of each layer obtained by wavelet decomposition j,k , the polarity direction is determined by the sign function sign(), and the phase characteristics of the original signal are retained;

[0081] The absolute value of the detail coefficient |d j,k ∣ Adaptive threshold λ to the current level jPerform real-time comparison and perform differential processing based on the comparison results. j,k ∣≥λ j When , the noise energy is reduced proportionally, when |d j,k ∣<λ j When , it is determined to be a valid signal component, the coefficient value is kept unchanged, and the signal distortion caused by traditional hard threshold processing is avoided;

[0082] The processed detail coefficients and the unprocessed approximate coefficients are subjected to inverse wavelet transform to reconstruct the denoised IMF components.

[0083] Furthermore, the expression of the bivariate model is:

[0084]

[0085] Where △F represents the temperature drift compensation of the tension measurement value, k1 and k2 represent the static temperature drift coefficient and dynamic temperature drift coefficient, respectively, T represents the current temperature of the tension sensor, T0 represents the ambient temperature, and dT / dt represents the temperature change rate.

[0086] Furthermore, the step of correcting the tension sensor data of the cantilever beam using the bivariate model includes:

[0087] Monitor bearing temperature in real time and activate dynamic compensation when the temperature change rate exceeds the preset threshold;

[0088] Correct zero drift according to the calibration curve.

[0089] Step 3: Construct an LSTM-Transformer dual-output network. The input is the preprocessed multi-source data, and the output includes two items. The first output item is the real-time tension value of the processed wire in a certain period of time in the future, and the second output item is the speed adjustment value of the stranding drive motor of the wire processing device and the pitch adjustment value of the stranding pitch adjustment mechanism.

[0090] Furthermore, the LSTM-Transformer dual-output network includes an input layer, two bidirectional LSTM layers, a multi-head attention layer and a fully connected layer. The input layer is used to receive a 28-dimensional feature vector, which can be set to a 5ms time window and contains vibration spectrum data, bearing temperature data, pulley inclination data, speed deviation data and 6-channel time-series tension sensor data.

[0091] Combine Figure 4 , where the workflow of the LSTM-Transformer dual-output network is:

[0092] (1) First, the input layer receives the raw data stream from the multimodal sensor, and the timing alignment module completes the multi-source data fusion. The specific process is as follows:

[0093] Data channel division: The input layer receives 28-dimensional feature vectors, which are divided into the following types according to data type: 6-channel tension time series data, with a sampling rate of 1kHz; 12-dimensional vibration spectrum data after FFT transformation, with a sampling rate of 2kHz; 2-dimensional pulley inclination data, with a sampling rate of 500Hz; 1-dimensional bearing temperature, with a sampling rate of 100Hz; 1-dimensional speed deviation, calculated by the encoder, with a sampling rate of 1kHz; 6-dimensional process parameters, i.e., preset values, including wire diameter, number of strands, etc.

[0094] Time series alignment mechanism: Linear interpolation is performed on low-frequency data such as bearing temperature to align it with the highest sampling rate, such as the vibration spectrum of 2kHz; data is cut into 5ms time windows to generate a tensor dimension of uniform time series length: 28×200, where 200 is the time step length.

[0095] (2) Then it is input to the bidirectional LSTM layer. In the first bidirectional LSTM layer, the number of hidden units is 128 in the forward direction and 128 in the reverse direction, totaling 256. It receives the aligned 28×200 time series data and outputs a 256-dimensional feature vector for each time step, with a dimension of 200×256. In the second bidirectional LSTM layer, the number of hidden units is 128 in the forward direction and 128 in the reverse direction. Its input is the 256-dimensional feature vector output by the first bidirectional LSTM layer, and its output is a 256-dimensional feature vector for each time step, with a dimension of 200×256.

[0096] (3) The outputs of the first-layer bidirectional LSTM and the second-layer bidirectional LSTM are concatenated along the feature dimension to form a 512-dimensional fusion feature. The dimension of the concatenated data is 200×512. The shallow and deep temporal features are fused through skip connections to improve the ability to capture sudden working conditions such as line disconnection transients.

[0097] (4) Input to the multi-head attention layer, convert the spliced ​​512-dimensional fusion features into query, key, and value matrices: generate Q / K / V of 8 attention heads through the linear projection layer, each head has a dimension of 64×200; perform self-attention calculation on each attention head, the formula is:

[0098]

[0099] Among them, d k =64,d k The dimensions for each head.

[0100] The outputs of the eight attention heads are concatenated into a 512-dimensional vector (8×64), and the dimensions are restored to 200×512, completing the multi-head output concatenation. This multi-head mechanism separates the attention weights of different modal features. For example, the weight of vibration features in sudden working conditions can be increased by 37%, enhancing model interpretability.

[0101] (5) Next, it is input to the dual-output fully connected layer, including: tension prediction branch (as output 1) and control parameter branch (as output 2). In the tension prediction branch, the 512-dimensional features output by the attention layer are converted to 256 dimensions through the fully connected layer 1, and the activation function used is ReLU. The 256-dimensional features are output as tension values ​​with a dimension of 1 through the fully connected layer 2, and the loss function used is the mean square error (MSE);

[0102] In the control parameter branch, the 512-dimensional features output by the attention layer are converted into 256-dimensional features through the fully connected layer 1, and the activation function used is LeakyReLU; the 256-dimensional features are converted into 2-dimensional features through the fully connected layer 2, including the speed adjustment amount Δn and the pitch adjustment amount Δp, and the loss function used is Huber loss.

[0103] The parameters of the previous network are shared by two branches, and the weight update paths of tension prediction and control parameters are separated during back propagation through gradient masking.

[0104] The speed deviation data is calculated using the following formula:

[0105] Δn=n set -n actual

[0106] Where n set is the preset twisting drive motor speed setting value, n actual The actual motor speed is measured in real time by the encoder. It is collected by the servo motor's built-in 17-bit absolute encoder with a sampling frequency of 1kHz.

[0107] Initialize the parameters of the LSTM-Transformer dual-output network and perform offline pre-training. Use the TD3 algorithm to optimize the control strategy online. The TD3 algorithm is a reinforcement learning algorithm for continuous control tasks.

[0108] Furthermore, the reward function of the LSTM-Transformer dual-output network is:

[0109]

[0110] Where σ2 is the standard deviation of tension, Δn is the speed adjustment, and Δp is the pitch adjustment.

[0111] In step 4, the output data of the LSTM-Transformer dual-output network is converted into a pulse width modulation signal to drive the servo motor and synchronize with the programmable logic controller to adjust the pitch hydraulic valve to control the pitch parameters of the wire stranding process.

[0112] In step 4, the control signal conversion and execution process specifically includes the following steps:

[0113] (1) Speed ​​control signal generation:

[0114] The speed adjustment value Δn output by the LSTM-Transformer network is converted into a pulse width modulation signal through a proportional mapping algorithm. The duty cycle calculation formula is:

[0115] D=K p ×Δn+D0

[0116] Where D is the final output duty cycle percentage, p = 0.15% / rpm is the proportional coefficient, determined by the servo motor speed-duty cycle calibration curve; D0 = 50% is the base duty cycle, corresponding to the mid-point of the motor's rated speed;

[0117] Output duty cycle limit range is 20% to 80% to prevent driver overload;

[0118] (2) Pitch hydraulic valve control:

[0119] The hydraulic valve opening control instruction is generated according to the pitch adjustment amount Δp, and the conversion formula is:

[0120]

[0121] Among them, V open It represents the percentage of hydraulic valve opening, and Pmax is the pitch adjustment corresponding to the maximum stroke of the hydraulic valve, which is calibrated based on the linear relationship between the hydraulic cylinder piston stroke and the pitch adjustment.

[0122] The duty cycle limiting algorithm avoids the risk of equipment damage under extreme operating conditions. The hydraulic valve opening calculation takes into account the nonlinear characteristics of the mechanical transmission system. The Pmax parameter is used to achieve a smooth transition from open-loop control to closed-loop control. In this example, Pmax can be set to 2.5mm.

[0123] (3) Real-time synchronous control:

[0124] Control instructions are transmitted to the programmable logic controller via the Modbus-TCP industrial communication protocol. Specific parameters are set as follows: synchronization period: 5ms, matching the fastest response requirements of the wire twisting process; synchronization period: 5ms, matching the fastest response requirements of the wire twisting process; timeout retransmission mechanism: triggers a system abnormality alarm after three consecutive communication failures.

[0125] Step 5: Check whether the current state meets the control conditions of the robust control mode. If yes, switch to the preset robust control mode and trigger the sound and light alarm.

[0126] Furthermore, the step of detecting whether the current state satisfies the control conditions of the robust control mode, and if so, switching to the preset robust control mode includes:

[0127] Establish control conditions, including:

[0128] The first condition is that the root mean square value of the triaxial vibration acceleration in any direction exceeds 5 times the acceleration of gravity and the duration is 200ms;

[0129] The second condition is that the bearing temperature is greater than 85°C or the temperature rise rate is greater than 2°C / min;

[0130] Detect the current state. If the first or second condition is met, switch the action, including:

[0131] The output items of the LSTM-Transformer dual-output network are frozen, and the servo motor is switched to a PID controller. The set value of the PID controller is initialized to the real-time speed measurement at the switching moment. The speed adjustment range is limited to ±1 rpm, and the pitch adjustment is frozen. A red alarm is triggered on the human-machine interface, and the main operation screen displays a red flashing border. A three-level alarm dialog box pops up, containing the fault code and occurrence time. The fault code is pushed to the manufacturing execution system, which automatically generates a maintenance work order and assigns a maintenance task.

[0132] This invention optimizes measurement deviations caused by environmental interference and structural limitations in traditional solutions through a dynamic compensation mechanism that integrates tension, angle, and vibration data. A multi-threaded processing architecture based on real-time control enables efficient conversion of sensor data into control instructions. The coordinated adjustment of the frequency converter and servo motor improves the ability to control the consistent tension of multiple strands. Furthermore, the invention integrates a hierarchical safety protection and maintenance mode design, simplifying the manual operation process and reducing reliance on professional experience, providing a viable technical path for high-precision and high-stability operation of wire processing equipment.

Claims

1. A multi-modal intelligent tension adaptive control method, characterized in that: The steps include: Acquire multi-source data from wire processing devices in real time, including tension sensor data, vibration time-domain signals, pulley inclination data, bearing temperature data, and process parameter data, and perform fast Fourier transform on the vibration time-domain signals to generate vibration spectrum data; Preprocessing of multi-source data includes: compensating tension sensor data; using an improved empirical mode decomposition algorithm to separate vibration noise and retain qualified IMF components; and correcting the cantilever beam tension sensor data using a bivariate model. Build an LSTM-Transformer dual-output network. The input is preprocessed multi-source data, and the output includes two items: the first output item is the real-time tension value of the processed wire within a certain period of time in the future, and the second output item is the speed adjustment value of the stranding drive motor of the wire processing device and the pitch adjustment value of the stranding pitch adjustment mechanism. The output data of the LSTM-Transformer dual-output network is converted into a pulse-width modulated signal to drive the servo motor and synchronized with the programmable logic controller to adjust the pitch hydraulic valve to control the pitch parameters of the wire stranding process. Detect whether the current state meets the control conditions of the robust control mode. If yes, switch to the preset robust control mode and trigger the sound and light alarm.

2. The multi-modal intelligent tension adaptive control method according to claim 1, characterized in that: The formula for compensating tension sensor data is: Where, F true Indicates the actual tension value after compensation, F raw represents the original tension value, M represents the weight of the pulley, θ represents the real-time pulley inclination angle, which is used to compensate for the measurement error when the pulley rotates, ω represents the speed of the capstan, r represents the distance from the center of the pulley to the center of the main shaft of the wire processing equipment, and φ represents the angle between the centrifugal force and the wire tension measurement direction.

3. The multi-modal intelligent tension adaptive control method according to claim 2, characterized in that: The steps of using the improved empirical mode decomposition algorithm to separate vibration noise and retain the qualified IMF components include: Determine the time delay δt between the vibration signal and the tension signal through cross-correlation analysis, and align the time axes of the vibration signal and the tension signal according to the time delay δt; Perform 5-layer empirical mode decomposition on the tension signal F(t) to obtain IMF1 to IMF5 components; Calculate the Pearson correlation coefficient r between each IMF component and the vibration signal a(t) i , mark|r i IMF components with |>0.7 are noise-dominated modes, where i=1,2,3,4,5; Wavelet soft thresholding is performed on the noise-dominated mode, and the threshold is calculated as Where σ1 is the noise standard deviation, N is the signal length; Retain Pearson correlation coefficient | r i |<0.5 IMF component, synthesize the tension signal F after noise reduction vib_comp .

4. The multi-modal intelligent tension adaptive control method according to claim 3, characterized in that: The structure of the LSTM-Transformer dual-output network includes an input layer, two bidirectional LSTM layers, a multi-head attention layer, and a fully connected layer. The input layer is used to receive a 28-dimensional feature vector, which contains 6-channel time-series tension sensing data, vibration spectrum data, temperature data, inclination data, and speed deviation data.

5. The multi-modal intelligent tension adaptive control method according to claim 4, characterized in that: The reward function of the LSTM-Transformer dual-output network is: R=10 / (1+σ2)-0.1∣Δn∣-0.05∣Δp∣, Where σ2 is the standard deviation of tension, Δn is the speed adjustment, and Δp is the pitch adjustment.

6. The multi-modal intelligent tension adaptive control method according to claim 5, characterized in that: The steps of detecting whether the current state satisfies the control conditions of the robust control mode and, if so, switching to the preset robust control mode include: Establish control conditions, including: The first condition is that the RMS value of the triaxial vibration acceleration in any direction exceeds 5g and the duration is 200ms; The second condition is that the bearing temperature is greater than 85°C or the temperature rise rate is greater than 2°C / min; Detect the current state. If the first or second condition is met, switch the action, including: The output items of the LSTM-Transformer dual-output network are frozen, and the PID controller is switched to control the servo motor. The speed adjustment range is limited to ±1 rpm, and the pitch adjustment is frozen. A red alarm is triggered on the human-machine interface, and the main operation screen displays a red flashing border. A three-level alarm dialog box pops up, containing the fault code and occurrence time. The fault code is then pushed to the manufacturing execution system, which automatically generates a maintenance work order and assigns a maintenance task.

7. The multi-modal intelligent tension adaptive control method according to claim 6, characterized in that: The expression of the bivariate model is: Where △F represents the temperature drift compensation of the tension measurement value, k1 and k2 represent the static temperature drift coefficient and dynamic temperature drift coefficient, respectively, T represents the current temperature of the tension sensor, T0 represents the ambient temperature, and dT / dt represents the temperature change rate.

8. The multi-modal intelligent tension adaptive control method according to claim 7, characterized in that: The steps of correcting the tension sensor data of the cantilever beam using the bivariate model include: Monitor bearing temperature in real time and activate dynamic compensation when the temperature change rate exceeds the preset threshold; Correct zero drift according to the calibration curve.

Images