Intelligent tightening curve state identification method and system

Abstract

The invention relates to an intelligent tightening curve state recognition method and system, belongs to the technical field of bolt assembly quality control, and aims at solving the problems that a traditional threshold method is high in omission ratio, a waveform analysis method is poor in generalization, and time sequence modeling of a machine learning model is insufficient. According to the technical scheme, torque signals and angle signals are collected in real time; constructing a dynamic observation sequence containing torque difference and torque angle gradient; defining five-stage hidden state sets S1 to S5; updating the state transition matrix through a sequential constraint optimization algorithm; decoding the optimal state sequence by adopting a viterbi algorithm; and triggering early warning based on the yield point probability threshold. The method has the technical effects that the abnormal state recognition precision is remarkably improved, the real-time detection capability is enhanced, the omission ratio is reduced, and dynamic working condition self-adaption is realized.

Description

Technical Field

[0001] This invention belongs to the field of bolt assembly quality control technology, and relates to an intelligent tightening curve state recognition method and system. Background Technology

[0002] Bolt tightening is a core assembly process in the automotive, aerospace, and precision equipment manufacturing industries, and its quality directly affects the safety and reliability of the overall structure. Traditional testing methods typically determine whether the tightening result is acceptable by setting fixed torque or angle thresholds. However, such methods can only capture static parameters at the endpoint and cannot dynamically identify abnormal states during the tightening process. For example, critical failure modes such as stripped threads, gasket deformation, and abnormal material yielding are easily missed under traditional threshold criteria.

[0003] In existing technologies, waveform analysis methods attempt to improve detection accuracy by manually extracting features such as slope change points of the tightening curve. However, this method is highly dependent on expert experience, sensitive to noise, and lacks generalization ability. On the other hand, machine learning-based methods, while able to learn curve features, face two major bottlenecks: first, they require a large amount of labeled data for model training, while abnormal working condition samples in industrial scenarios are scarce; second, existing models do not adequately model the time-series dynamic features, making it difficult to capture the state transition process of bolts from elastic deformation to plastic yielding.

[0004] The aforementioned technical deficiencies lead to three core problems:

[0005] (1) Traditional methods require waiting for the tightening to finish before the result can be judged, and cannot provide real-time warnings of abnormalities during the process;

[0006] (2) The static threshold method is not sensitive to process anomalies, resulting in a long-term failure rate of more than 15%;

[0007] (3) The existing system cannot dynamically adjust the detection logic according to material properties and working conditions, and frequent manual calibration is required.

[0008] Therefore, there is an urgent need for a state recognition solution that can analyze the dynamic characteristics of the tightening process in real time, has online learning capabilities, and is highly noise resistant, in order to meet the requirements of zero defects in assembly quality in the high-end manufacturing field. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide an intelligent tightening curve state recognition method and system, which solves the technical problems in the prior art such as weak real-time detection of bolt tightening curves, missed detections, and inaccuracy of large amounts of data standards.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for intelligently identifying tightening curve states includes the following steps:

[0012] S1: Real-time acquisition of torque signal T during the tightening process q (t) and the angle signal θ(t), where t represents time;

[0013] S2: Constructing a dynamic observation sequence O t =[T q (t),ΔT(t),G TA [(t),θ(t)], where ΔT(t) is the torque difference, G TA (t) represents the torque-angle gradient;

[0014] S3: Define the hidden state set S = {S1, S2, S3, S4, S5}, where S1 represents the idling stage, S2 represents the elastic deformation stage, S3 represents the plastic deformation initiation stage, S4 represents the complete yielding stage, and S5 represents the slippage fracture stage.

[0015] S4: Iteratively update the state transition matrix A = [a] using a sequential constraint optimization algorithm. ij ] 5×5 The updated formula is: Where λ=e -kΔt Here, Δt is the time decay factor, and Δt is the update time interval. Let be the count of the transition from state i to state j at time t;

[0016] S5: Decode the optimal state sequence using the Viterbi algorithm. The formula for calculating path probability is: δ t (j)=max i [δ t-1 (i)·a ij ]·b j (O t ), where b j (O t ) represents the observation probability;

[0017] S6: When The time marker is the seating point, when The point where P(S4) > 0.95 is marked as the yield point.

[0018] Furthermore, in S2, the torque-angle gradient G TA The calculation of (t) uses the Savitzky-Golay filtering differentiation method:

[0019]

[0020] Where Δθ = 0.1° is the angular interval, m = 4 is the window half-width, and c k These are the differential convolution kernel coefficients.

[0021] Furthermore, in step S4, the initialization of the state transition matrix A is obtained through historical data statistics:

[0022]

[0023] Where N ij This represents the frequency of transitions from state i to state j in the training set. The training set contains no fewer than 2000 sets of tightening curve data, and the initial probability distribution is π = [0.95, 0.05, 0, 0, 0]. T .

[0024] Furthermore, in S5, the observation probability b j (O t Calculated using a Gaussian Mixture Model (GMM):

[0025]

[0026] Where M = 4 is the number of mixture components, ω jm For the mixed weights to satisfy μ jm Let ∑ be the mean vector. jm The covariance matrix is ​​obtained by training historical data using the EM algorithm.

[0027] Furthermore, in step S4, the update process includes dynamically optimizing the parameters of the Hidden Markov Model (HMM) using the Baum-Welch algorithm, with the objective function being:

[0028]

[0029] Where λ represents the HMM parameter set λ = (π, A, B).

[0030] Furthermore, the process before step S1 includes a preprocessing step: friction compensation is applied to the torque signal, and the compensation model is as follows:

[0031] T true =T means -T friction

[0032]

[0033] Where T means For measuring torque, μ0 is the basic coefficient of friction, μ1 is the transient coefficient of friction, and θ is the tightening angle. For the characteristic angle, F axial This is the axial preload; and the angle signal is smoothed using cubic spline interpolation.

[0034] θ(t)=a i +b i (t-t1)+c i (t-t1) 2 +d i (t-t1) 3

[0035] where t∈[t i ,t i+1 ], coefficient a i b i c i d i By using the boundary conditions θ′0(0)=b0=0 and θ″ n-1 (t end =0, which confirms the result.

[0036] A tightening status monitoring system for implementing the method includes:

[0037] Torque sensor with a range of 0 to 1000 N·m and a linearity error ≤ ±0.3% FS;

[0038] Photoelectric encoder with angular resolution ≤0.01° and response frequency ≥3kHz;

[0039] Embedded processing unit with built-in Hidden Markov Model (HMM) inference engine module;

[0040] The torque sensor and photoelectric encoder are connected to the embedded processing unit via a bus to transmit signals in real time; the embedded processing unit executes:

[0041] (i) Calculate the observation vector O every 10ms. t ;

[0042] (ii) Update the state transition matrix A every 100ms;

[0043] (iii) When P(S4) > 0.95, output the yield point alarm signal, or when P(S5) > 0.5, output the safety range over-limit signal and cut off the power.

[0044] Furthermore, the embedded processing unit employs a Field-Programmable Gate Array (FPGA) to implement Viterbi parallel decoding, and the hardware architecture includes:

[0045] The feature extraction module calculates ΔT(t) and G in real time. TA (t);

[0046] The probability calculation unit integrates a parallel multiplier to calculate δ. t (j);

[0047] The path backtracking memory uses a first-in-first-out (FIFO) queue with a depth of ≥5k;

[0048] The feature extraction module and the probability calculation unit are interconnected via a data bus, and the path backtracking memory is connected to the output of the probability calculation unit.

[0049] Furthermore, the tightening status monitoring system also includes a retraining trigger module, which triggers a retraining process when certain conditions are met. and When this occurs, the Baum-Welch algorithm is triggered to retrain the HMM parameters.

[0050] The beneficial effects of this invention are as follows:

[0051] (1) By using a hidden Markov model to perform time-series modeling of the tightening process, the torque signal, torque difference, torque angle gradient, and angle signal are fused into a dynamic observation sequence, enabling for the first time to track the entire process of the bolt from the idle stage, elastic deformation stage, plastic deformation initiation stage, complete yielding stage to stripping and fracture stage. This mechanism fundamentally solves the defect of the traditional threshold method, which only focuses on the endpoint torque and ignores the dynamic characteristics of the process, so that progressive anomalies such as stripping and gasket deformation can be accurately captured in the initial stage.

[0052] (2) A torque angle gradient calculation based on Savitzky-Golay filtering is introduced to effectively suppress noise interference in industrial settings while preserving key features of curvature changes. The state transition matrix is ​​dynamically updated using a sequential constraint optimization algorithm, and the probability distribution is corrected in real time using a time decay factor, enabling the model to adapt to complex working conditions such as different material properties and friction coefficient variations. The optimal state path is decoded based on the Viterbi algorithm, and the observation probability is calculated using a Gaussian mixture model to ensure that yield point identification simultaneously meets both probability significance requirements and physical state constraints.

[0053] (3) The embedded system uses FPGA to implement parallel Viterbi decoding. Through a three-stage pipeline architecture consisting of a feature extraction module, a probability calculation unit, and a path backtracking memory, the state recognition delay is compressed to the millisecond level. The dual buffering mechanism ensures the timing coordination of signal acquisition, feature extraction, and model update, meeting the data throughput requirements of high-frequency response sensors. When the probability of the complete yielding stage exceeds the set threshold or the safety range exceeds the limit, the system can trigger power-off protection within a millisecond response time, significantly improving the safety level of the equipment.

[0054] (4) The multi-stage state recognition capability advances the abnormal interception point to the plastic deformation stage, avoiding batch assembly defects; the dynamic learning mechanism reduces the frequency of manual calibration and lowers maintenance costs; the retraining trigger module can autonomously judge the model mismatch condition, providing self-evolution capability for the intelligent manufacturing system.

[0055] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0057] Figure 1 This is a system structure diagram of the present invention;

[0058] Figure 2 This is a flowchart of the data sampling process for the double buffering mechanism of the present invention;

[0059] Figure 3 This is a flowchart of the conflict detection retraining process of the present invention;

[0060] Figure 4 The figure shows the experimental results of this invention. Detailed Implementation

[0061] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0062] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0063] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0064] A method for intelligently identifying tightening curve states includes the following steps:

[0065] Dynamic feature extraction and observation sequence establishment, defining a four-dimensional observation vector O t =[T q (t),ΔT(t),G TA [(t),θ(t)], where the torque signal T q The signals θ(t) and angle signals θ(t) are derived from real-time historical data acquisition at a sampling frequency of 1 kHz. For torque differential, This represents the torque-angle gradient.

[0066] To improve the noise resistance of the torque-angle gradient data, Savitzky-Golay filtering and differentiation are used to optimize the data. Where c k The coefficients are the differential convolution kernel coefficients, m = 4, Δθ = 0.1°; Table 1 is the parameter table of the differential convolution kernel coefficients.

[0067] Table 1

[0068] k ±4 ±3 ±2 ±1 0 <![CDATA[c k ]]> ±0.018 ±0.092 ±0.139 ±0.315 0

[0069] The HMM state space defines the tightening stage, with a state set S = {S1, S2, S3, S4, S5}, where S1 is the idling stage, S2 is the elastic deformation stage, S3 is the plastic deformation initiation stage, S4 is the complete yielding stage, and S5 is the stripping fracture stage. Table 2 shows the observation characteristics corresponding to the five physical stages.

[0070] Table 2

[0071] state Physical stage Observational features <![CDATA[s1]]> Idle phase <![CDATA[T q (t)<50N·m,G TA (t)≈0]]> <![CDATA[s2]]> Elastic deformation stage <![CDATA[G TA (t)∈[0.1,0.5]N·m / °]]> <![CDATA[s3]]> Initial stage of plastic deformation ΔT = [1.5, 2] × ΔT(t-1) <![CDATA[s4]]> Complete yielding stage <![CDATA[G TA (t)<0.1N·m / °]]> <![CDATA[s5]]> Slipped tooth fracture stage <![CDATA[T q (t)=(1±0.2)T q (t-1)]]>

[0072] Establish a probability model and set the initial probability distribution π = [0.95, 0.05, 0, 0, 0]. T That is, there is a 95% probability of starting in state S1. Let the initial state transition matrix A be:

[0073]

[0074] Setting the observation probability model Right now:

[0075]

[0076] Where M = 4, μ jm Let ∑ be the mean vector. jm Let μ be the covariance matrix. jm and ∑ jm All data were trained using 2000 sets of historical data, with mixed weights ω. jm satisfy

[0077] After the model is built, the Baum-Welch algorithm is used to maximize the likelihood function:

[0078]

[0079] Sequence constraint optimization is performed, and a time decay factor λ = e is introduced. -kΔt Let Δt = 0.02, meaning the matrix is ​​updated every 50ms, and λ = e -0.01k The transition probability is updated dynamically, and the probability formula is:

[0080]

[0081] Real-time state decoding is performed recursively using the Viterbi algorithm, initializing δ1(i) = π. i b i (O1), ψ1(i) = 0 are used for recursive calculation:

[0082] δ t (j)=max i [δ t-1 (i)·a ij ]·b j (O t )

[0083] ψ t (j)=arg{max i [δ t-1 (i)·a ij ]}

[0084] The optimal probability and the optimal hidden state at the final time step are expressed as follows:

[0085] P * =max i δ T (i)

[0086]

[0087] Then, based on the path backtracking formula, the optimal state sequence can be derived in reverse:

[0088]

[0089] like Figure 1 As shown in the overall structure diagram, the hardware configuration includes a torque sensor, an angle encoder, and an embedded processor.

[0090] like Figure 2 As shown, the real-time data stream management adopts a double buffering mechanism, with a signal acquisition period of 1ms, a feature extraction period of 10ms, and an HMM update period of 50ms.

[0091] HMM parameter initialization and training: Data acquisition is standardized, requiring hexagonal head bolts of M6 to M20, materials ranging from carbon steel, stainless steel, and titanium alloy, with an abnormal working condition stripping rate ≥15%, washer deformation rate ≥10%, and sampling frequency of 1kHz. 2000 sets of tightening curves (including normal and abnormal states) are collected, and five-state transition points are manually labeled to form a labeled dataset. The state transition frequency is statistically analyzed based on... Initialize the transition matrix A0, and initialize the observation probability B0 using a Gaussian mixture model (GMM):

[0092]

[0093] The parameters were obtained through optimization using the EM algorithm.

[0094] Online dynamic optimization using the Baum-Welch algorithm to maximize the likelihood function:

[0095]

[0096] The goal is to find a set of parameters λ such that the observation sequence O0, O1, ..., O T The probability of occurrence L(λ) reaches its maximum.

[0097] Real-time state recognition process: First, the torque signal is corrected, and a friction compensation model is proposed: T true =T means -T friction

[0098] A friction coefficient model dependent on angle is adopted:

[0099]

[0100] Where θ is the tightening angle, μ0 is the basic friction coefficient, μ1 is the transient friction coefficient, and F... axial It is the axial preload. The characteristic angle.

[0101] Then, the angle is interpolated using an interpolation algorithm to ensure smooth endpoints.

[0102] θ(t)=a i +b i (t-t1)+c i (t-t1) 2 +d i (t-t1) 3

[0103] Where, t∈[t i ,t i+1 ] represents the location of the difference point, t1 is a given node, and a i b i c i d i The coefficients are undetermined and are determined by the following conditions.

[0104] (1) Interpolation conditions:

[0105] θ(t i )=y i

[0106] θ(t i+1 ) = a i +b i (t i+1 -t i )+c i (t i+1 -t i ) 2 +d i (t i+1 -t i ) 3 =y i+1

[0107] (2) The first derivative is continuous:

[0108] θ′ i (t i+1 )=θ′ i+1 (t i+1 )

[0109] (3) The second derivative is continuous:

[0110] θ i (t) i+1 )=θ i " +1 (t i+1 )

[0111] (4) Boundary conditions:

[0112] θ′0(0)=b0=0

[0113] θ n " -1 (t end )=b end-1 +2c emd-1 (t end -t end-1 )+3d emd-1 (t end -t end-1 ) 2 =0

[0114] After the original data preprocessing is completed, feature extraction is performed to establish new feature gradients. Then, a Savitzky-Golay differentiator is used to fit the torque sequence. Where c k Here are the differential convolution kernel coefficients, m = 4, Δθ = 0.1°. This yields the complete observation sequence O. t =[T q (t),ΔT(t),G TA (t),θ(t)].

[0115] To prevent underflow, a logarithmic field transformation is used to compress the probability to (-∞, 0).

[0116] logδ t (j)=max i [δ t-1 (i)+loga ij +logb j (O t )

[0117] By using a path backtracking compression algorithm to store state change points, the space complexity is reduced from O(T) to O(1) compared to the traditional method of recording the state at every time t. This achieves the goal of only recording the jump points (t, q). t ).

[0118] like Figure 3 As shown, when a state transition conflict occurs, a retraining condition is triggered, and a mathematical judgment expression for the trigger is established:

[0119] and

[0120] in This represents the transition probability of state i to state j in the tk-th training iteration. This represents the posterior probability of the transition at the current time t.

[0121] like Figure 4As shown in Table 3, compared with traditional methods (traditional threshold method and window analysis method), the innovative technology has significantly improved the rework rate and scrap rate in terms of landing point recognition rate, yield point delay, online self-adaptation, economic benefits, and economic efficiency.

[0122] Table 3

[0123]

[0124] Example 1: Dynamic Observation Sequence Construction and State Recognition Process

[0125] Step 11: Signal Acquisition and Preprocessing

[0126] A high-precision torque sensor acquires the raw torque signal T at a frequency of 1kHz. means The photoelectric encoder synchronously acquires the angle signal θ with a resolution of 0.01°;

[0127] Perform friction compensation:

[0128]

[0129] Where μ0 = 0.12 is the basic friction coefficient of carbon steel, and μ1 = 0.05 is the transient attenuation factor;

[0130] The angle signal is smoothed using cubic spline interpolation, and the boundary conditions satisfy the condition that the derivatives at the start and end points are zero.

[0131] Step 12: Feature Extraction and Observation Sequence Generation

[0132] Calculate the torque difference:

[0133] The torque angle gradient is calculated using the Savitzky-Golay filter differentiation: convolution kernel coefficients c k Take the values ​​from Table 1;

[0134] Construct a four-dimensional observation vector O t =[T q (t),ΔT(t),G TA [(t),θ(t)];

[0135] The observation sequence is output to the HMM inference engine every 10ms.

[0136] Example 2: Online Optimization Process for Hidden Markov Models

[0137] Step 21: Model Initialization

[0138] Loading pre-trained parameters:

[0139] The initial state distribution is π = [0.95, 0.05, 0, 0, 0]. T;

[0140] The state transition matrix A0 was obtained based on statistics from 2000 sets of historical data;

[0141] The observation probability B0 is a Gaussian mixture model with four components;

[0142] Configure time decay factor λ = e -0.01k Update cycle 50ms;

[0143] Step 22: Real-time parameter optimization

[0144] The forward probability α is calculated using the Baum-Welch algorithm. t (i) and backward probability β t (i) Statistical state transition count

[0145] Update the transition matrix in sequence:

[0146] Retraining is triggered when an abnormal transfer pattern is detected:

[0147] There exists a state transition pair i→j that satisfies the sum of historical transition probabilities being less than 0.03 and the current...

[0148] Example 3: FPGA Parallel Decoding and Decision Process

[0149] Step 31: Hardware Architecture Execution

[0150] The feature extraction module adopts a pipeline design: the first stage calculates ΔT(t), and the second stage calculates G. TA (t);

[0151] The probability computation unit deploys five parallel multipliers: synchronously calculating δ t (j)=max i [δ t-1 (i)·a ij ]·b j (O t );

[0152] The path backtracking memory uses a FIFO queue with a depth of 5120.

[0153] Only store the state transition points t and q. t ;

[0154] Step 32: Security Decision-Making Mechanism

[0155] Mark the location when three consecutive decoding states are S3;

[0156] A yellow warning signal is triggered when P(S4) > 0.95;

[0157] When P(S5) > 0.5, a red emergency power-off signal is output;

[0158] The tightening tool can be stopped suddenly via the GPIO interface.

[0159] Finally, 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 present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligently recognizing the state of a tightening curve, characterized in that: Includes the following steps: S1: Real-time acquisition of torque signal T during the tightening process q (t) and the angle signal θ(t), where t represents time; S2: Constructing a dynamic observation sequence O t =[T q (t),ΔT(t),G TA [(t),θ(t)], where ΔT(t) is the torque difference, G TA (t) represents the torque-angle gradient; S3: Define the hidden state set S = {S1, S2, S3, S4, S5}, where S1 represents the idling stage, S2 represents the elastic deformation stage, S3 represents the plastic deformation initiation stage, S4 represents the complete yielding stage, and S5 represents the slippage fracture stage. S4: Iteratively update the state transition matrix A = [a] using a sequential constraint optimization algorithm. ij ] 5×5 The updated formula is: Where λ=e -kΔt Here, Δt is the time decay factor, and Δt is the update time interval. Let be the count of the transition from state i to state j at time t; S5: Decoding the optimal state sequence using the Viterbi algorithm. The formula for calculating path probability is: δ t (j)=max i [δ t-1 (i)·a ij ]·b j (O t ), where b j (O t ) represents the observation probability; S6: When The time marker is the seating point, when The point where P(S4) > 0.95 is marked as the yield point.

2. The intelligent tightening curve state recognition method according to claim 1, characterized in that: In S2, the torque-angle gradient G TA The calculation of (t) uses the Savitzky-Golay filtering differentiation method: Where Δθ = 0.1° is the angular interval, m = 4 is the window half-width, and c k These are the differential convolution kernel coefficients.

3. The intelligent tightening curve state recognition method according to claim 1, characterized in that: In step S4, the initialization of the state transition matrix A is obtained through historical data statistics: Where N ij This represents the frequency of transitions from state i to state j in the training set. The training set contains no fewer than 2000 sets of tightening curve data, and the initial probability distribution is π = [0.95, 0.05, 0, 0, 0]. T .

4. The intelligent tightening curve state recognition method according to claim 1, characterized in that: In S5, the observation probability b j (O t ) Calculated using Gaussian Mixture Model (GMM): Where M = 4 is the number of mixture components, ω jm For the mixed weights to satisfy μ jm Let ∑ be the mean vector. jm The covariance matrix is ​​obtained by training historical data using the EM algorithm.

5. The intelligent tightening curve state recognition method according to claim 1, characterized in that: In step S4, the update process includes dynamically optimizing the parameters of the Hidden Markov Model (HMM) using the Baum-Welch algorithm, with the objective function being: Where λ represents the HMM parameter set λ = (π, A, B).

6. The intelligent tightening curve state recognition method according to claim 1, characterized in that: The process before S1 includes a preprocessing step: friction compensation of the torque signal, with the compensation model as follows: T true =T means -T friction Where T means For measuring torque, μ0 is the basic coefficient of friction, μ1 is the transient coefficient of friction, and θ is the tightening angle. For the characteristic angle, F axial This is the axial preload; and the angle signal is smoothed using cubic spline interpolation. θ(t)=a i +b i (t-t1)+c i (t-t1) 2 +d i (t-t1) 3 where t∈[t i ,t i+1 ], coefficient a i b i c i d i By using the boundary conditions θ0'(0)=b0=0 and θ″ n-1 (t end =0, which confirms the result.

7. A tightening status monitoring system for implementing the method of any one of claims 1 to 6, characterized in that: include: Torque sensor with a range of 0 to 1000 N·m and a linearity error ≤ ±0.3% FS; Photoelectric encoder with angular resolution ≤0.01° and response frequency ≥3kHz; Embedded processing unit with built-in Hidden Markov Model (HMM) inference engine module; The torque sensor and photoelectric encoder are connected to the embedded processing unit via a bus to transmit signals in real time; the embedded processing unit executes: (i) Calculate the observation vector O every 10ms. t ; (ii) Update the state transition matrix A every 100ms; (iii) When P(S4) > 0.95, output the yield point alarm signal, or when P(S5) > 0.5, output the safety range over-limit signal and cut off the power.

8. The tightening status monitoring system according to claim 7, characterized in that: The embedded processing unit uses a field-programmable gate array (FPGA) to implement Viterbi parallel decoding. The hardware architecture includes: The feature extraction module calculates ΔT(t) and G in real time. TA (t); The probability calculation unit integrates a parallel multiplier to calculate δ. t (j); The path backtracking memory uses a first-in-first-out queue with a depth of ≥5k. The feature extraction module and the probability calculation unit are interconnected via a data bus, and the path backtracking memory is connected to the output of the probability calculation unit.

9. The tightening status monitoring system according to claim 7, characterized in that: The tightening status monitoring system also includes a retraining trigger module, which triggers a retraining mechanism when certain conditions are met. and When this occurs, the Baum-Welch algorithm is triggered to retrain the HMM parameters.

Images