Method and system for actively preventing diamond wire of multi-wire cutting machine from being broken

Abstract

The invention discloses a multi-wire cutting machine diamond wire active breakage prevention method and system. The method comprises the following steps that a detection assembly used for collecting various feature data of a diamond wire is installed on a to-be-detected multi-wire cutting machine; obtaining multiple groups of historical multi-class feature data, and performing fusion processing on the multiple groups of historical multi-class feature data to obtain multiple groups of fused historical feature vector sequences; constructing a mixed deep learning model, and training the mixed deep learning model by using the plurality of groups of historical feature vector sequences to obtain a prediction model; acquiring a feature vector sequence of the to-be-detected multi-wire cutting machine in real time and inputting the feature vector sequence into the prediction model; and taking corresponding active prevention measures based on the prediction result and the judgment threshold. According to the method, the tension, the vibration, the temperature, the speed and the arch height of the diamond wire are comprehensively considered, the running state of the diamond wire can be reflected more comprehensively and accurately, the accuracy of wire breakage risk prediction is improved, and corresponding active prevention measures can be taken after the wire breakage risk is predicted.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of diamond wire cutting, and particularly relates to a method and system for actively preventing breakage of diamond wire of a multi-wire cutting machine. BACKGROUND

[0002] As a key device for cutting hard and brittle materials (such as silicon rods) in the photovoltaic and semiconductor industries, the multi-wire cutting machine uses high-speed moving diamond wire to perform friction cutting on the material through the cutting blade attached to the steel wire. With the development of the industry, the requirements for cutting efficiency and precision are continuously increasing, and the wire diameter of the diamond wire is continuously refined. However, the reduction in wire diameter makes the diamond wire more susceptible to breakage during cutting due to various factors.

[0003] In actual cutting process, the diamond wire needs to be in a high-speed moving state for a long time and bear a large tension and friction. For example, in the photovoltaic industry, in order to improve production efficiency, the cutting speed of the multi-wire cutting machine is continuously increased, which results in a significant increase in mechanical stress on the diamond wire. At the same time, the heat generated during cutting also affects the performance of the diamond wire, such as causing a decrease in hardness and toughness, thereby increasing the risk of breakage.

[0004] In order to prevent the diamond wire from breaking, some devices in the prior art use a method of manually checking the diamond wire at regular intervals. This method is not only inefficient, but also difficult to detect potential damage in real time and prevent breakage accidents in a timely manner due to the subjectivity and limitations of manual detection. Some devices in the prior art only rely on a tension detection device to monitor the tightness of the diamond wire. When the tension exceeds a preset threshold, an alarm is issued. However, this method has obvious limitations. It cannot predict failures in advance, and when the tension fluctuates greatly, false alarms may occur, thereby interfering with the normal production process. Moreover, the causes of diamond wire breakage in actual production are complex and diverse, and it is difficult to comprehensively and effectively prevent wire breakage by relying on a single detection means. SUMMARY

[0005] The first object of the present application is to provide a multi-wire cutting machine diamond wire active breakage prevention method that can comprehensively consider multiple factors and has high accuracy.

[0006] The second object of the present application is to provide a multi-wire cutting machine diamond wire active breakage prevention system.

[0007] Technical solution: The present application discloses a multi-wire cutting machine diamond wire active breakage prevention method, comprising the following steps,

[0008] S1: installing a detection assembly for collecting multiple types of feature data of the diamond wire on the multi-wire cutting machine to be detected;

[0009] S2: Obtain multiple sets of historical multi-class feature data, respectively fuse the multiple sets of historical multi-class feature data, and obtain multiple sets of fused historical feature vector sequences X;

[0010] S3: Construct a hybrid deep learning model combining LSTM and attention mechanism, train the hybrid deep learning model using the multiple sets of historical feature vector sequences X, and obtain a prediction model for predicting the diamond wire fracture probability based on the feature vector sequence X;

[0011] S4: Real-time acquisition of multi-class feature data of the multi-wire cutting machine to be detected, fusion processing of the multi-class feature data, input of the feature vector sequence X obtained after the fusion processing into the prediction model, and output of the prediction result R by the prediction model;

[0012] S5: Based on the prediction result R and the predetermined judgment threshold R threshold , take appropriate proactive prevention measures; the proactive prevention measures include: when R≥R threshold , issue an alarm and shut down the multi-wire cutting machine; when R threshold , adjust the diamond wire based on the prediction result R, return to step S4, and shut down the multi-wire cutting machine.

[0013] Further, the detection assembly in step S1 includes a tension sensor for collecting tension data of the diamond wire, a vibration sensor for collecting vibration data of the multi-wire cutting machine, a temperature sensor for collecting temperature data of the cutting area, a line speed sensor for collecting running speed data of the diamond wire, and a laser displacement sensor for collecting arch height data of the diamond wire.

[0014] Further, the fusion processing of the multi-class feature data in step S2 is as follows:

[0015] Preprocess the collected multi-class feature data;

[0016] Extract the feature parameters of each type of feature data after preprocessing to obtain the original feature vector X i (t) of each type of feature data;

[0017] Standardize the original feature vector X i (t) of the multi-class feature data to obtain the target feature vector

[0018] Calculate the initial weight w i (0) of the target feature vector of each type of feature data using the analytic hierarchy process AHP to obtain the initial weight vector w(0);

[0019] updating the target feature vector of each type of feature data based on the initial weight vector w(0) and the actual contribution of each type of feature data weight;

[0020] updating the target feature vector of each type of feature data based on the initial weight vector w(0) and the actual contribution of each type of feature data weight, to a plurality of target feature vectors performing fusion to obtain an initial comprehensive feature vector F total (t);

[0021] performing sliding window optimization processing on the initial comprehensive feature vector F total (t) to obtain a target comprehensive feature vector F based on the target comprehensive feature vector F calculating to obtain a feature vector sequence X.

[0022] Further, the feature parameters of each type of preprocessed feature data are extracted as follows:

[0023] extracting instantaneous tension, tension rate of change and average tension from the tension data to form the original feature vector X1(t) of the first type of feature data;

[0024] extracting vibration frequency and amplitude from the vibration data to form the original feature vector X2(t) of the second type of feature data;

[0025] extracting instantaneous temperature, temperature rate of change and average temperature from the temperature data to form the original feature vector X3(t) of the third type of feature data;

[0026] extracting instantaneous linear speed and its differential quantity, as well as tension disturbance features indirectly mapped by speed fluctuations from the running speed data to form the original feature vector X4(t) of the fourth type of feature data;

[0027] extracting the linear bow feature vector from the bow height data to form the original feature vector X5(t) of the fifth type of feature data.

[0028] Further, the target feature vector of each type of feature data is updated in the following manner: weight,

[0029] the actual contribution of each type of feature data is defined as the feature contribution degree C i (t),

[0030]

[0031] wherein γ i is the feature change sensitivity coefficient of the i-th type of feature data;

[0032] The updating weight formula is:

[0033] w i (t) = λ · w i (t-1) + (1 - λ) · C i (t)

[0034] wherein t-1 represents the last time; λ is a weight smoothing coefficient, and 0<λ<1;

[0035] wherein the characteristic change sensitivity coefficient γ i is calculated as follows:

[0036] The absolute deviation sum of the real-time state vector and the reference state vector is calculated, and the calculation formula is as follows:

[0037]

[0038] wherein Z(t) is the real-time state vector, Z(t)=[Z1(t),Z2(t),Z3(t),Z4(t),Z5(t)], Z1(t) corresponds to the original feature vector X1(t) of the tension data, Z2(t) corresponds to the original feature vector X2(t) of the vibration data, Z3(t) corresponds to the original feature vector X3(t) of the temperature data, Z4(t) corresponds to the original feature vector X4(t) of the running speed data, and Z5(t) corresponds to the original feature vector X5(t) of the bow height data; Z ref =[Z 1,ref ,Z 2,ref ,...,Z 5,ref ]is the reference vector under normal working conditions, and Z 1,ref is the mean value of the original feature vector X1(t), Z 2,ref is the mean value of the original feature vector X2(t), Z 3,ref is the mean value of the original feature vector X3(t), Z 4,ref is the mean value of the original feature vector X4(t), and Z 5,ref is the mean value of the original feature vector X5(t); ΔZ(t) is the Manhattan distance between Z(t) and Z ref , and ΔZ(t) is used to represent the absolute deviation sum of the real-time state vector and the reference state vector;

[0039] The change amount of each type of feature data is calculated, and for the standardized features of the i-th type of feature data, the change amount is

[0040] The correlation strength between ρ i and ΔZ(t) is quantified by Pearson correlation coefficient: i

[0041] After normalization, ρ i is multiplied by the initial value γi,init fusion, to obtain final gamma i:

[0042] Further, wherein the initial comprehensive feature vector F total (t) is obtained in the following manner:

[0043]

[0044] Wherein theta (t) is a compensation term, and the calculation formula of theta (t) is as follows:

[0045]

[0046] Wherein kappa is a correlation parameter, and 0 < kappa < 1; cov(X i ,X j ) is the covariance of the i-th feature data and the j-th feature data.

[0047] Further, the target comprehensive feature vector F is obtained in the following manner:

[0048]

[0049] Wherein N is the size of the sliding window;

[0050] The feature vector sequence X is calculated based on the target comprehensive feature vector F in the following manner:

[0051] The feature data is continuously collected with T as the time step, and the target comprehensive feature vector F at all discrete time points is calculated. t T The final feature vector sequence X = [x1, x2,..., x t ,..., x ] is obtained by arranging the target comprehensive feature vectors F t in chronological order; each element x T in the feature vector sequence X is a 5-dimensional feature vector, and the overall sequence length is T.

[0052] Further, the manner of constructing a hybrid deep learning model by combining LSTM and attention mechanism in step S3 is as follows:

[0053] The feature vector sequence X is converted into the format required by the input layer of the LSTM in the prediction model;

[0054] The LSTM performs time series feature extraction, and the LSTM network processes the time series dependency relationship through the forget gate, input gate and output gate to generate a hidden state sequence H = [h1, h2,..., h t ,..., h T ];

[0055] The attention mechanism acts on the hidden state sequence H = [h1, h2,..., hT] of the LSTM output. The attention mechanism highlights key time sequence features by calculating time step weights, and the calculation is as follows: t ,...,h T ] The attention mechanism highlights key time sequence features by calculating time step weights, and the calculation is as follows:

[0056] The ability value e t is calculated to measure the importance of each hidden state, and the calculation formula of the ability value e t is as follows: t e t = a (h T , h a ) = W h · tanh (W t h H + W T h a + b t ) ;

[0057] Then, the e t is normalized by softmax to obtain the attention weight a t ,

[0058] The context vector c is generated to fuse the hidden state

[0059] wherein, T is the time step, h t is the output hidden state of the LSTM at the t time step; e t ∈R m is the energy value of the hidden state at the t time; a (·) is an attention score function, and an additive attention score function is adopted; a t ∈ [0, 1] is the attention weight at the t time; c ∈ R m is the context vector fused with key features; Wa ∈ R 1×m , W h ∈R m×n , W H ∈R m×n are weight matrices of the attention mechanism, respectively; b a ∈R m is a bias vector of the attention mechanism.

[0060] The full connection layer and the output layer process, the context vector c is input into the full connection layer for feature conversion, and finally the prediction result R of the diamond wire breaking risk is obtained through the output layer.

[0061] Further, the way of adjusting the diamond wire based on the prediction result R in step S5 is as follows:

[0062] The maximum allowable tension T max, the calculation formula is as follows:

[0063]

[0064] Wherein σ b is the tensile strength of the diamond wire, in N / m 2 ; d is the diameter of the diamond wire, in m; k s is the safety factor;

[0065] Based on the prediction result R, the target tension T of the diamond wire is calculated target , the calculation formula is as follows:

[0066] T target =T base ·(1-α·R)

[0067] T min ≤T target ≤T max

[0068] Wherein, T min is the relaxation threshold, in N; T base is the basic tension, in N; α is the risk adjustment coefficient, 0<α<1;

[0069] Based on the current tension and the target tension of the diamond wire, the tension adjustment value ΔT is calculated, and the calculation formula of ΔT is as follows:

[0070] ΔT=T target -T current

[0071] Wherein T current is the current tension, T current is monitored in real time by the tension sensor;

[0072] Based on the tension adjustment value ΔT, the tension adjustment speed v T is calculated, and the calculation formula of the tension adjustment speed v T is as follows:

[0073]

[0074] v T ≤v T,max

[0075] Wherein, v T is the tension adjustment speed, in N / s; γ is the smoothing coefficient; t reg is the target adjustment time, in s; v T,max is the maximum adjustment speed, in N / s;

[0076] Based on the tension adjustment value ΔT and the tension adjustment speed v TThe tension adjustment value ΔT is corrected in real time, and the motor of the multi-wire cutting machine is controlled based on a PID control method, and the calculation formula for controlling the motor is as follows:

[0077]

[0078] Wherein, u(t) is a motor control signal, used to drive the motor to adjust the tension; K p is a proportional gain of the PID control, K i is an integral gain of the PID control, K d is a differential gain of the PID control.

[0079] Based on the same inventive concept, the application also discloses a multi-wire cutting machine diamond wire active anti-breaking system, comprising,

[0080] a feature data acquisition module, used to acquire multiple types of feature data of the diamond wire;

[0081] a data fusion processing module, used to fuse the acquired multiple types of feature data to obtain a fused feature vector sequence X;

[0082] a prediction model training module, which combines LSTM and attention mechanism to construct a hybrid deep learning model, obtains multiple groups of historical feature vector sequences X based on the feature data acquisition module and the data fusion processing module, and trains the hybrid deep learning model by using the multiple groups of historical feature vector sequences X to obtain a prediction model for predicting the breaking probability of the diamond wire based on the feature vector sequence X;

[0083] a breaking risk prediction module, which can obtain the feature vector sequence X of the multi-wire cutting machine to be detected in real time based on the feature data acquisition module and the data fusion processing module, and input the feature vector sequence X into the prediction model, and the prediction model outputs a prediction result R;

[0084] a motor control execution module, which can obtain the prediction result R and a preset judgment threshold R threshold take corresponding active prevention measures, shut down the multi-wire cutting machine or control the motor of the multi-wire cutting machine to run; when the active prevention measures taken are to control the motor of the multi-wire cutting machine to run, the breaking risk prediction module and the motor control execution module are cyclically executed until the multi-wire cutting machine is shut down;

[0085] a man-machine interaction module, used to output and display the multiple types of feature data of the diamond wire, the prediction result R and the corresponding active prevention measures, and realize man-machine interaction.

[0086] Beneficial effects: Compared with the prior art, the present application has the following remarkable advantages: the present application comprehensively considers the tension, vibration, temperature, speed and bow height of the diamond wire, can more comprehensively and accurately reflect the running state of the diamond wire relative to a single detection means, and improves the accuracy of the broken wire risk prediction; and after predicting the broken wire risk, the present application takes corresponding active prevention measures based on the prediction result, immediately closes the equipment to reduce the loss when the risk is high, and adjusts the diamond wire to the optimal running state when the risk is low, which meets the actual production demand and prevents the broken wire. The present application combines LSTM and attention mechanism to construct a hybrid deep learning model, the hybrid deep learning model can extract the time sequence features for accurate screening and strengthening of key information without losing the time dimension dependence, which is beneficial to improve the prediction accuracy of the prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1 is a flowchart of the method of the present application;

[0088] Figure 2 is a structural schematic diagram of the system of the present application. DETAILED DESCRIPTION

[0089] The technical solutions of the present application will be further described below in combination with the drawings.

[0090] Example 1

[0091] The multi-wire saw diamond wire active broken wire prevention method disclosed in this embodiment, as shown in Figure 1 , includes the following steps,

[0092] S1: installing a detection assembly for collecting multi-type feature data of the diamond wire on the multi-wire saw to be detected.

[0093] The multi-type feature data includes the following 5 types of feature data: the first type of feature data is the tension data of the diamond wire; the second type of feature data is the vibration data of the multi-wire saw; the third type of feature data is the temperature data of the cutting area; the fourth type of feature data is the running speed data of the diamond wire; and the fifth type of feature data is the bow height data of the diamond wire.

[0094] The detection assembly comprises a tension sensor, a vibration sensor, a temperature sensor, a linear velocity sensor and a laser displacement sensor, wherein the tension sensor is installed on a diamond wire tensioning device of the multi-wire cutting machine and is used to collect tension data of the diamond wire in real time; the vibration sensor is installed on a cutting device of the multi-wire cutting machine and is used to collect vibration data in real time, preferably, is installed on key parts of the cutting device, such as a cutting motor shell, a cutting workbench and the like; the temperature sensor is installed on a contact area of the diamond wire and the cutting material and is used to collect temperature data, and the temperature sensor is preferably an infrared temperature sensor; the linear velocity sensor is installed on a driving wheel of the diamond wire and is used to collect running speed data; and the laser displacement sensor is installed above a diamond wire cutting area and is used to collect bow height data, the laser displacement sensor measures the bow height of the diamond wire by emitting a laser beam and receiving reflected light, is used to monitor the bow variable of the wire and can accurately obtain the change of the wire bow.

[0095] S2: Obtain a plurality of sets of historical multi-class feature data, and perform fusion processing on the plurality of sets of historical multi-class feature data respectively to obtain a plurality of sets of fused historical feature vector sequences X.

[0096] S21: Preprocess the collected multi-class feature data, wherein the preprocessing includes filtering, denoising processing, removing high-frequency noise and baseline drift interference. Preprocessing the feature data can improve the data quality thereof.

[0097] S22: Extract feature parameters of each type of feature data after preprocessing to obtain an original feature vector X of each type of feature data. i (t).

[0098] The manner of extracting the feature parameters is as follows:

[0099] extract instantaneous tension, tension change rate and average tension from the tension data to constitute an original feature vector X1(t) of the first type of feature data, wherein F inst (t) is the instantaneous tension at time t; is the tension change rate, Δt1 is the sampling interval of the tension data; F avg (t) is the average tension, N1 is the size of the sliding sampling window of the tension data.

[0100] extract vibration frequency and amplitude from the vibration data to constitute an original feature vector X2(t) of the second type of feature data, X2(t) = [f dom (t), A peak (t)] T , wherein f dom(t) is the amplitude of the vibration data. peak (t) is the amplitude of the vibration data.

[0101] The instantaneous temperature, the temperature change rate and the average temperature are extracted from the temperature data to form the original feature vector X3(t) of the third type of feature data, where Tem inst (t) is the instantaneous temperature at time t; is the temperature change rate, Δt3 is the sampling interval of the temperature data; Tem avg (t) is the average temperature, N3 is the size of the sliding sampling window of the temperature data.

[0102] The instantaneous linear speed and its differential quantity, and the tension disturbance feature indirectly mapped by the speed fluctuation are extracted from the running speed data to form the original feature vector X4(t) of the fourth type of feature data, v inst (t) is the instantaneous linear speed at time t; is the speed differential quantity, that is, the acceleration, Δt4 is the sampling interval of the running speed data; ΔF dist (t) is the tension disturbance feature, which is the tension anomaly mapped by the speed fluctuation through dynamic coupling, and the calculation steps are as follows:

[0103] 1) The deviation of the instantaneous linear speed from the reference linear speed is defined as the speed fluctuation quantity Δv(t) at time t, and the calculation formula of the speed fluctuation quantity Δv(t) is Δv(t) = v inst (t) - v base (t), where v base (t) is the reference linear speed, which is the average speed in the sliding sampling window: N4 is the size of the sliding sampling window of the running speed data;

[0104] 2) The historical running speed data under the normal working condition of the diamond wire is obtained, and the linear relationship between the tension change and the speed fluctuation is fitted, and the fitted linear relationship is ΔF exp (t) = k·Δv(t) + b, where k is the coupling coefficient, b is the bias term, and k and b are solved by the least square method; ΔF exp (t) is the fitted tension change;

[0105] 3) The deviation of the measured tension from the expected tension (under the normal state of the diamond wire) is defined as the tension disturbance feature ΔF dist (t), and the calculation formula of the tension disturbance feature ΔF dist (t) is ΔFdist (t) = |ΔF(t) - ΔF exp (t) |, ΔF dist (t) is used to represent the disturbance intensity, the greater the value indicates the tension deviates from the normal mode more significantly, where ΔF(t) = F inst (t) - F base (t), F base (t) is the reference tension, i.e. the expected tension at normal speed, F inst (t) is the instantaneous tension at time t.

[0106] The line bow feature vector is extracted from the bow height data to form the original feature vector X5(t) of the fifth type of feature data, X5(t) = [S(t)] T , where S(t) is the line bow value at time t, i.e. the maximum distance of the diamond line from the reference straight line.

[0107] S23: standardizing the original feature vectors X i (t) of the multi-type feature data to obtain the target feature vectors X

[0108] Before standardizing the original feature vectors of the multi-type feature data, the time scale and sampling frequency of the original feature vectors of the multi-type feature data are unified, and t0 obtained after unifying the time scale and sampling frequency is the sampling start time, t end is the sampling start time, and Δt is the sampling interval, i.e. t ∈ [t0, t end ]. Different sensors have different sampling frequencies, which are unified by linear interpolation.

[0109] Preferably, before standardizing the original feature vectors of the multi-type feature data, the original feature vectors of the multi-type feature data are subjected to outlier rejection.

[0110] The original feature vectors X i (t) of the multi-type feature data are standardized in the following manner:

[0111]

[0112] , where X i (t) represents the original feature vector of the i-th type of feature data, i = 1, 2, 3, 4, 5; is the target feature vector of the i-th type of feature data after standardization, μ i is the mean of the original feature vector of the i-th type of feature data, and σ i is the standard deviation of the original feature vector of the i-th type of feature data. The final five groups of target feature vectors are Data standardization processing can eliminate the influence of dimension difference between different original features, and is beneficial to subsequent fusion processing.

[0113] S24: The initial weight w of the target feature vector of each type of feature data is calculated by using analytic hierarchy process (AHP) i (0), to obtain the initial weight vector w(0), w(0)=[w1(0), w2(0), w3(0), w4(0), w5(0)] T .

[0114] The initial weight w i (0) is calculated according to the following formula:

[0115]

[0116] w i (0) is the initial weight of the target feature of the i-th type of feature data, and alpha i is a feature importance coefficient, and the value range is 0-1, alpha i has a physical meaning based on the standardized feature, the tension directly affects the cutting accuracy, the vibration reflects the stability of the equipment, alpha i needs to be determined based on domain knowledge, historical fault data and experimental verification, such as tension alpha1=0.25, vibration alpha2=0.2, temperature alpha3=0.15, running speed alpha4=0.2, and bow height alpha5=0.2; beta i is a signal-to-noise ratio of the target feature vector, which reflects the reliability of the i-th target feature vector, the higher the signal-to-noise ratio, the more reliable the data, beta i needs to be calculated based on statistical analysis of sensor data, beta i = signal energy / noise energy.

[0117] The initial weight distribution of the target feature vector in the application considers both the physical importance of the feature and the data quality characteristics, so that the initial weight is more reasonable.

[0118] S25: Based on the initial weight vector w(0) and the actual contribution of each type of feature data, the target feature vector of each type of feature data is updated weight.

[0119] The target feature vector of each type of feature data is updated weight in the following manner:

[0120] The actual contribution of each type of feature data is defined as the feature contribution degree C i (t),

[0121]

[0122] Where gamma iis the feature change sensitivity coefficient of the ith type of feature data, and γ i ∈ [0, 1], γ i The larger the value is, the more likely it is that a slight change in the type of feature data will cause a significant fluctuation in the operating state, such as the risk of line breakage caused by a sudden change in tension being much greater than the risk of line breakage caused by a slow change in temperature.

[0123] wherein the feature change sensitivity coefficient γ i is calculated as follows:

[0124] (1) Calculate the absolute deviation sum of the real-time state vector and the reference state vector, and the calculation formula is as follows:

[0125]

[0126] wherein Z(t) is the real-time state vector, Z(t) = [Z1(t), Z2(t), Z3(t), Z4(t), Z5(t)] = [X1(t), X2(t), X3(t), X4(t), X5(t)], Z1(t) corresponds to the original feature vector X1(t) of the tension data, Z2(t) corresponds to the original feature vector X2(t) of the vibration data, Z3(t) corresponds to the original feature vector X3(t) of the temperature data, Z4(t) corresponds to the original feature vector X4(t) of the operating speed data, and Z5(t) corresponds to the original feature vector X5(t) of the bow height data; Z ref = [Z 1,ref , Z 2,ref ,..., Z 5,ref ] is the reference vector under normal operating conditions, and Z 1,ref is the mean value of the original feature vector X1(t), Z 2,ref is the mean value of the original feature vector X2(t), Z 3,ref is the mean value of the original feature vector X3(t), Z 4,ref is the mean value of the original feature vector X4(t), and Z 5,ref is the mean value of the original feature vector X5(t); ΔZ(t) is the Manhattan distance between Z(t) and Z ref , and ΔZ(t) is used to represent the absolute deviation sum of the real-time state vector and the reference state vector;

[0127] (2) Calculate the change amount of each type of feature data. For the standardized features of the ith type of feature data, the change amount is

[0128] (3) Use the Pearson correlation coefficient to quantify the correlation strength between γ and ΔZ(t):

[0129] (4) Calculate the feature change sensitivity coefficient γ iNormalized sensitivity coefficient of feature variation γ i Initial value γ i,init Fusion, get the final γi:

[0130] Wherein, the weight 0.7 and 0.3 balance the initial value set by physical mechanism and data association respectively, and in actual application, it can be adjusted according to application scene and data characteristics, the core principle is to balance the influence weight of priori knowledge and data driven, and ensure γ i ∈[0,1].

[0131] The update weight formula is:

[0132] w i (t)=λ·w i (t-1)+(1-λ)·C i (t)

[0133] Wherein t-1 represents the last time, λ is a weight smoothing coefficient (0<λ<1), preferably 0.6≤λ≤0.8; the greater λ is, the more gentle the weight change is, avoiding high frequency fluctuation; the smaller λ is, the more sensitive the weight response to the current contribution is, suitable for the scene of rapid change of working condition, used to balance the influence of historical weight and current contribution. This mechanism enables the weight to automatically adjust with the change of running state, improving the adaptability of the algorithm to complex working conditions.

[0134] S26: based on the updated target feature vector Weight, fuse multiple target feature vectors Get the initial comprehensive feature vector F total (t).

[0135] Wherein, the way to calculate the initial comprehensive feature vector F total (t) is as follows:

[0136]

[0137] Wherein θ(t) is a compensation term, and the calculation formula of θ(t) is as follows:

[0138]

[0139] Wherein κ is a correlation parameter, and 0<κ<1; cov(X i ,X j ) is the covariance of the i-th type of feature data and the j-th type of feature data. The introduction of compensation term can effectively utilize the correlation information between different types of features, and improve the accuracy of fusion result.

[0140] S27: sliding window optimization processing is performed on the initial comprehensive feature vector F total (t), and the target comprehensive feature vector Based on target comprehensive feature vector The eigenvector sequence X is calculated.

[0141] Where the initial comprehensive feature vector F total The optimization formula for sliding window optimization (t) is as follows:

[0142]

[0143] Where N is the sliding window size, which is determined based on the sampling frequency. This step can further improve the stability and reliability of the fusion results, providing high-quality feature input for subsequent diamond wire operation status evaluation.

[0144] For continuously acquired data, the smoothed target comprehensive feature vector is calculated for each discrete time point t (t = 1, 2, ..., T, where T is the total time step, determined by the sampling frequency and effective data length of the original data) using an optimized formula. Each of them It is a 5-dimensional vector.

[0145] The calculated target comprehensive feature vectors at all discrete time points Arrange them sequentially according to time to form a temporal feature sequence, resulting in the final feature vector sequence X = [x1, x2, ..., x...]. t ,...,x T ].

[0146] Let the smooth eigenvector at time t=1 The first element x1 of the corresponding feature vector sequence X, i.e.

[0147] Let the smooth eigenvector at time t=2 The second element x2 corresponding to the feature vector sequence X, i.e. ...

[0149] Let the smooth eigenvector at time t = T The T-th element x of the corresponding feature vector sequence X T ,Right now

[0150] The target integrated feature vector with discrete time points After concatenation in chronological order, the feature vector sequence X is obtained: Each element x in the feature vector sequence X t All are 5-dimensional feature vectors, and the overall sequence length is T, i.e., the time step is T.

[0151] The present application comprehensively considers the five kinds of information of tension, vibration, temperature, speed and wire bow height, wherein the tension directly reflects the stress condition of the diamond wire and is closely related to the fracture risk and cutting accuracy; the vibration reflects the stability of the equipment and the diamond wire during operation and has an influence on the fatigue wear rate; the temperature reflects the thermal environment of the cutting area and is related to the performance degradation and material metamorphic risk of the diamond wire; the speed determines the cutting efficiency and load balance; and the wire bow height shows the deformation of the diamond wire in shape and is related to the cutting contact state and tension anomaly. The present application comprehensively considers the five kinds of information, which can comprehensively cover the working state of the diamond wire from different core dimensions, and the five kinds of information work together to accurately identify potential problems from multiple dimensions of stress, motion, environment and shape, thereby realizing comprehensive and accurate judgment of the state of the diamond wire. The detection result is more reliable than the single detection means in the prior art, and the limitation of single dimension monitoring in the prior art is compensated.

[0152] S3: combining LSTM and attention mechanism to construct a hybrid deep learning model, and training the hybrid deep learning model by using a plurality of historical feature vector sequences X to obtain a prediction model for predicting the fracture probability of the diamond wire based on the feature vector sequence X.

[0153] The fracture condition of the diamond wire in each historical feature vector sequence is known, and the historical feature vector sequence X is labeled according to the actual situation. The labeling can be performed in the following manner: 1 represents broken wire, and 0 represents unbroken wire.

[0154] The way of combining LSTM and attention mechanism to construct a hybrid deep learning model is to apply the attention mechanism to the output layer of the LSTM, specifically to the hidden state sequence H=[h1, h2,...,h t ,...,h T ] output by the LSTM. The attention mechanism highlights key time sequence features by calculating time step weights.

[0155] Since the feature vector sequence X needs to meet the format requirement of the time sequence feature vector sequence for the input layer of the prediction model (LSTM), the feature vector sequence X needs to be converted into the format required by the input layer of the LSTM in the prediction model, that is, the input layer of the prediction model (LSTM) linearly transforms the feature vector sequence X to adapt it to the input dimension of the LSTM network. The adaptation conversion formula is:

[0156]

[0157] Wherein, is the input vector at the tth time after adaptation; W t is the input vector at the tth time before adaptation; and W in ∈R m×n , Win Let m be the weight matrix of the input layer, m be the dimension of the LSTM hidden layer, and n be the weight matrix of x. t Dimensions; b in ∈R m b in This is the bias vector for the input layer.

[0158] LSTM for temporal feature extraction: The LSTM network processes temporal dependencies through forget gates, input gates, and output gates to generate a hidden state sequence H = [h1, h2, ..., h t ,...,h T The specific calculations are as follows:

[0159] (1) The forgetting gate determines the proportion of historical information discarded:

[0160] (2) The input gate updates the cell state with new information:

[0161] (3) Cell state renewal:

[0162] (4) Output Gate and Hidden State: Output Gate and Hidden State: h t =o t ⊙tanh(C t );

[0163] Among them, f t ∈R m f t i is the output vector of the forget gate; t ∈R m i t The output vector of the input gate; o t ∈R m o t C is the output vector of the output gate; t ∈R m Let be the cell state at time t; These represent the candidate cell states at time t; h t ∈R m Let W be the hidden state at time t; f W i W C W o ∈R m×(m+n) W represents the weight matrices for the forget gate, input gate, and output gate of the LSTM, respectively. f W is the weight matrix of the forget gate; i W C The input gate part corresponds to two weight matrices, where Wi is the weight matrix of the input gate for calculating the input gate signal i(t), W C is the weight matrix of the candidate state for calculating the candidate cell state; W o is the weight matrix of the output gate; b f ,b i ,b C ,b o ∈R m are the bias vectors of the forget gate, the input gate and the output gate of the LSTM, respectively, b f is the bias vector of the forget gate; b i ,b C are the two bias vectors corresponding to the input gate part, where b i is the bias vector of the input gate, which is attributed to the calculation of the input gate, used to adjust the activation threshold of the sigmoid function, thereby affecting the default opening degree of the gate; b C is the bias vector of the candidate state, which is attributed to the calculation of the candidate state, used to adjust the output center of the tanh function, thereby affecting the default value of the generated new information; b o is the bias vector of the output gate; σ is the sigmoid activation function; and is the element-level multiplication (Hadamard product).

[0164] The attention mechanism is directed to the hidden state sequence H of the LSTM output, H = [h1, h2,..., h t ,...,h T ], and the attention mechanism highlights the key time sequence features by calculating the time step weight, in the following manner:

[0165] (1) Energy value e t is calculated: the energy value is used to measure the importance of each hidden state, and the calculation formula of the energy value e t is e t = a(h t ,h T ) = W a · tanh(W h h t +W H h T +b a );

[0166] (2) Then, the e t is subjected to softmax normalization to obtain the attention weight a t ,

[0167] (3) Context vector c is generated: the weighted fusion hidden state

[0168] wherein, T is the time step, h t is the output hidden state of LSTM at the t time step; e t ∈R m is the energy value of the t time step hidden state; a(·) is an attention score function, and an additive attention score function is adopted; α t ∈[0,1] is the attention weight at the t time step; c∈R m is the context vector fused with key features; Wa∈R 1×m , W h ∈R m×n , W H ∈R m×n are weight matrices of the attention mechanism, respectively; b a ∈R m is the bias vector of the attention mechanism.

[0169] The full connection layer and the output layer process, the context vector c is input into the full connection layer for feature conversion, and finally the prediction result of the diamond wire breaking risk is obtained through the output layer. The output layer adopts a sigmoid activation function, and the prediction formula is:

[0170] y = σ(W o c + b o )

[0171] wherein y is the breaking risk prediction value, and the value range is y∈[0,1]; W o is the output layer weight matrix, b o is the output layer bias term, and σ is the sigmoid function.

[0172] In summary, the attention mechanism takes the hidden state sequence H = [h1, h2,..., h t ,...,h T ] output by LSTM as input, calculates the energy value e t and the weight α t of each hidden state h t , and realizes the differentiated weighting of the time sequence features. The core logic is: in the diamond wire breaking prediction task, the importance of features at different time steps is different, for example, the “tension surge” feature 10 seconds before breaking is more critical than the smooth feature during normal operation, and the attention mechanism automatically identifies the key time steps and gives them higher weights α t , strengthens the fusion of key features, and weakens the influence of irrelevant or noise features.

[0173] Progressive fusion logic of LSTM and attention mechanism: LSTM first completes the comprehensive extraction of the time sequence characteristics of the feature vector sequence, ensuring that the time dimension dependency is not lost, and then the attention mechanism further screens and strengthens the key information of the extracted time sequence characteristics, forming a collaborative process of first comprehensive capture and then focused attention. This fusion not only retains the dynamic characteristics of the time sequence characteristics, but also solves the limitation of LSTM treating all time step features "equally", preventing the final hidden state h T Generally only contains end information, ignoring the key moment.

[0174] In the risk prediction of diamond wire breakage, the combination of LSTM and attention mechanism can effectively capture long-term temporal dependencies and extract dynamic evolution features comprehensively. At the same time, the attention mechanism focuses on key warning moments, thereby reducing the influence of noise and redundant information. This significantly improves the representation ability, sensitivity of key signals and anti-interference ability of the hybrid deep learning model to fault rules, and ultimately improves the accuracy and robustness of the breakage risk prediction, providing high-quality decision support for the running state evaluation.

[0175] Preferably, after training the prediction model, a plurality of historical feature vector sequences with labeled diamond wire breakage are obtained and used as a validation set. The validation set is used to iteratively optimize the prediction model. The cross-entropy loss function is used to calculate the loss between the predicted value output by the prediction model and the actual label labeled in advance in the validation set. The Adam optimizer is used to optimize the parameters of the prediction model, and the prediction parameters are updated to minimize the cross-entropy loss function. The calculation formula of the cross-entropy loss function is as follows:

[0176]

[0177] Where M is the number of samples in the validation set, y i is the actual label labeled in advance in the i-th sample in the validation set, is the predicted value of the i-th sample.

[0178] S4: Real-time acquisition of multi-line cutting machine multi-class feature data, fusion processing of multi-class feature data, input of the feature vector sequence X obtained after fusion processing into the prediction model, and output of the prediction result R by the prediction model. The fusion processing method of the multi-class feature data is consistent with step S2.

[0179] S5: Based on the prediction result R and the judgment threshold R threshold Take appropriate proactive prevention measures; proactive prevention measures include: when R≥R threshold , issue an alarm and shut down the multi-line cutting machine; when R thresholdWhen the prediction result R is obtained, the diamond wire is adjusted based on the prediction result R, and the step S4 is returned until the multi-wire cutting machine is stopped.

[0180] wherein 0≤R threshold ≤1. Preferably, R threshold is set to 0.8; when R≥0.8, it is determined as a high risk state of wire breakage, and when R<0.8, it is determined as a low risk state of wire breakage.

[0181] When the wire breakage risk is determined as a high risk, an alarm is issued and the multi-wire cutting machine is stopped, which can effectively avoid serious damage to the equipment and the material being cut caused by the diamond wire breakage, and can reduce economic losses. When the wire breakage is determined as a low risk state, the motor of the multi-wire cutting machine is controlled based on the PID control method, and then the tension of the diamond wire is adjusted to adjust the stress of the diamond wire, so that the diamond wire is in the best operating state to prevent wire breakage.

[0182] The way of adjusting the diamond wire based on the prediction result R is as follows:

[0183] The maximum allowable tension T max of the diamond wire is calculated, and the calculation formula is as follows:

[0184]

[0185] T max is the maximum allowable tension of the diamond wire, and the unit is N, that is, the upper limit of the tension in the safety range; σ b is the tensile strength of the diamond wire, and the unit is N / m 2 ; d is the diameter of the diamond wire, and the unit is m; k s is a safety factor, which is usually taken as 0.6-0.8, and the setting of the safety factor can avoid the tension close to the material limit strength and reserve a safety margin;

[0186] The target tension T target of the diamond wire is calculated based on the prediction result R, and the calculation formula is as follows:

[0187] T target =T base ·(1-α·R)

[0188] T min ≤T target ≤T max

[0189] wherein, T target is the target tension, and the unit is N, that is, the ideal tension value to which the diamond wire needs to be adjusted; T min is a relaxation threshold, that is, the minimum safe tension allowed by the diamond wire system, and the unit is N; T base is a basic tension, and the unit is N, which is determined by process experiments, and R thresholdThe running tension of the diamond wire, R threshold is preferably 0.8; a is a risk adjustment coefficient, 0

[0190] The tension adjustment value ΔT is calculated based on the current tension and the target tension of the diamond wire, i.e. the difference between the current tension and the target tension:

[0191] ΔT = T target - T current

[0192] wherein ΔT is the tension adjustment value, in N, and ΔT is positive indicating that the tension needs to be increased, and ΔT is negative indicating that the tension needs to be decreased; T current is the current tension, T current is monitored in real time by the tension sensor;

[0193] The tension adjustment speed v T is calculated based on the tension adjustment value ΔT, and the tension adjustment speed v T is set to ensure stability during the adjustment of the tension of the diamond wire and to avoid impact of the diamond wire due to rapid adjustment. The calculation formula of the tension adjustment speed v T is as follows:

[0194]

[0195] v T ≤ v T,max

[0196] wherein v T is the tension adjustment speed, in N / s, i.e. the amount of change of the tension per unit time; γ is a smoothing coefficient, and γ is preferably 0.8-1.2, and γ is used to ensure smooth adjustment and reduce vibration of the diamond wire; t reg is the target adjustment time, in s, which is set according to the response speed of the multi-wire cutting machine, and t reg is preferably 0.5-2; v T,max is the maximum adjustment speed, in N / s, and t reg is preferably 0.2·T max , and t reg is positively correlated with the cutting speed.

[0197] The tension adjustment value ΔT is corrected in real time based on the tension adjustment value ΔT and the tension adjustment speed v T , and the motor of the multi-wire cutting machine is controlled based on the PID control method. The calculation formula of the control of the motor is as follows:

[0198]

[0199] Wherein, u(t) is a motor control signal (such as voltage or speed instruction) for driving the motor to adjust the tension; ΔT(t) = T target -T current (t), that is, the difference between the target tension and the current tension; K p is the proportional gain of the PID control, K i is the integral gain of the PID control, K d is the differential gain of the PID control.

[0200] Preferably, the embodiment further comprises S6: real-time output display of the multi-class feature data of the diamond wire, the prediction result R and the corresponding active prevention measures.

[0201] In actual application, a user-friendly visual monitoring interface can be developed by using the Qt framework, a dashboard is built by using the QtWidgets or QtQuick components, and key parameters such as the tension data of the diamond wire, the vibration data of the multi-wire cutting machine, the temperature data of the cutting area, the running speed data of the diamond wire, and the bow height data of the diamond wire are displayed in the form of numerical values, curves, and indicator lights in real time, and the data refresh frequency (1-10 Hz) is configurable.

[0202] In actual application, the working state data and the equipment running parameters of the diamond wire can be uploaded to a manufacturing execution system (MES) through a specified or specific data transmission interface, which is beneficial to providing data support for enterprise production decision-making.

[0203] Embodiment 2

[0204] The active anti-breaking system for the diamond wire of the multi-wire cutting machine disclosed in the present application comprises a feature data acquisition module, a data fusion processing module, a prediction model training module, a breaking risk prediction module, a motor control execution module, and a human-computer interaction module. Figure 2

[0205] The feature data acquisition module is used for acquiring the multi-class feature data of the diamond wire.

[0206] The multi-class feature data comprises the following five types of feature data: the first type of feature data is the tension data of the diamond wire; the second type of feature data is the vibration data of the multi-wire cutting machine; the third type of feature data is the temperature data of the cutting area; the fourth type of feature data is the running speed data of the diamond wire; and the fifth type of feature data is the bow height data of the diamond wire.

[0207] ​The tension sensor is installed on the diamond wire tensioning device of the multi-wire cutting machine to collect the tension data of the diamond wire in real time; the vibration sensor is installed on the cutting device of the multi-wire cutting machine to collect the vibration data in real time, preferably on the key parts of the cutting device, such as the shell of the cutting motor and the cutting workbench; the temperature sensor is installed on the contact area between the diamond wire and the cutting material to collect the temperature data, and the temperature sensor is preferably an infrared temperature sensor; the linear velocity sensor is installed on the driving wheel of the diamond wire to collect the running speed data; the laser displacement sensor is installed above the diamond wire cutting area to collect the bow height data, and the laser displacement sensor measures the bow height of the diamond wire by emitting a laser beam and receiving reflected light, which is used to monitor the bow variable of the wire and accurately obtain the change of the wire bow.

[0208] The data fusion processing module is configured to fuse the collected multi-type feature data to obtain a fused feature vector sequence X. The data fusion processing module performs step S2 in embodiment 1.

[0209] The prediction model training module is configured to construct a hybrid deep learning model combining LSTM and attention mechanism, obtain a plurality of historical feature vector sequences X based on the feature data acquisition module and the data fusion processing module, and train the hybrid deep learning model using the plurality of historical feature vector sequences X to obtain a prediction model for predicting the diamond wire fracture probability based on the feature vector sequence X. The prediction model training module performs step S3 in embodiment 1.

[0210] The fracture risk prediction module is configured to obtain the feature vector sequence X of the multi-wire cutting machine to be detected in real time based on the feature data acquisition module and the data fusion processing module, and input the feature vector sequence X into the prediction model. The prediction model outputs a prediction result R.

[0211] The motor control execution module is configured to control the motor of the multi-wire cutting machine based on the prediction result R and a predetermined judgment threshold R threshold Take appropriate proactive prevention measures to shut down the multi-wire cutting machine or control the motor operation of the multi-wire cutting machine; when the proactive prevention measure taken is to control the motor operation of the multi-wire cutting machine, the fracture risk prediction module and the motor control execution module are executed in a loop until the multi-wire cutting machine is shut down. The motor control execution module performs step S5 in embodiment 1.

[0212] The human-computer interaction module is configured to output and display the multi-type feature data of the diamond wire, the prediction result R and the corresponding proactive prevention measures, and realize human-computer interaction.

[0213] In practical application, a user-friendly visual monitoring interface can be developed using the Qt framework, and a dashboard can be built using Qt Widgets or Qt Quick components to display key parameters such as the tension data of the diamond wire, the vibration data of the multi-wire cutting machine, the temperature data of the cutting area, the running speed data of the diamond wire, and the bow height data of the diamond wire in the form of numerical values, curves, and indicator lights in real time. The data refresh frequency (1-10 Hz) can be configured.

[0214] In practical application, the working state data of the diamond wire and the equipment operating parameters can be uploaded to a manufacturing execution system (MES) through a specified or specific data transmission interface, which is beneficial to providing data support for enterprise production decisions.

Claims

1. A multi-wire saw diamond wire active anti-breaking method, characterized in that: The method comprises the following steps, S1: installing a detection assembly for collecting multi-class feature data of the diamond wire on the multi-wire saw to be detected; S2: obtaining a plurality of groups of historical multi-class feature data, respectively performing fusion processing on the plurality of groups of historical multi-class feature data, and obtaining a plurality of groups of fused historical feature vector sequences X; S3: combining LSTM and an attention mechanism to construct a hybrid deep learning model, and training the hybrid deep learning model by using the plurality of groups of historical feature vector sequences X to obtain a prediction model for predicting the breaking probability of the diamond wire based on the feature vector sequence X; S4: obtaining multi-class feature data of the multi-wire saw to be detected in real time, performing fusion processing on the multi-class feature data, inputting the feature vector sequence X obtained after the fusion processing into the prediction model, and outputting a prediction result R by the prediction model; S5: based on the prediction result R and a judgment threshold R set in advance threshold Take appropriate proactive countermeasures; The active prevention measures include: when R≥R threshold , issuing an alarm and shutting down the multi-wire saw; when R threshold , adjusting the diamond wire based on the prediction result R, and returning to step S4 until the multi-wire saw is shut down.

2. The multi-wire saw diamond wire active anti-breaking method according to claim 1, characterized in that: The detection assembly in step S1 comprises a tension sensor for collecting tension data of the diamond wire as the first-class feature data, a vibration sensor for collecting vibration data of the multi-wire saw as the second-class feature data, a temperature sensor for collecting temperature data of the cutting area as the third-class feature data, a wire speed sensor for collecting running speed data of the diamond wire as the fourth-class feature data, and a laser displacement sensor for collecting the bow height data of the diamond wire as the fifth-class feature data.

3. The multi-wire saw diamond wire active anti-breaking method according to claim 1, characterized in that: The fusion processing manner of the multi-class feature data in steps S2 and S4 is as follows: The collected multi-class feature data is preprocessed; The feature parameters of the preprocessed feature data of each type are extracted respectively to obtain the original feature vectors X of the feature data of each type i (t); The original feature vector X of the multi-class feature data i (t) standardization processing is performed to obtain the target feature vector of each class of feature data An analytic hierarchy process (AHP) is used to calculate a target feature vector of each type of feature data of the initial weight w i (0), to obtain an initial weight vector w(0); updating the target feature vector of each type of feature data based on the initial weight vector w(0) and actual contribution of each type of feature data weight based on the updated target feature vector weights, to the plurality of sets of target feature vectors to obtain an initial comprehensive feature vector F total (t); to the initial comprehensive feature vector F total (t) performing sliding window optimization processing to obtain the target comprehensive feature vector based on the target comprehensive feature vector The feature vector sequence X is calculated.

4. The multi-wire saw diamond wire active anti-breaking method according to claim 3, characterized in that: The feature parameters of the preprocessed multi-class feature data are extracted in the following manner: The instantaneous tension, tension change rate and average tension are extracted from the tension data to constitute the original feature vector X1(t) of the first-class feature data; The vibration frequency and amplitude are extracted from the vibration data to constitute the original feature vector X2(t) of the second-class feature data; The instantaneous temperature, temperature change rate and average temperature are extracted from the temperature data to constitute the original feature vector X3(t) of the third-class feature data; The instantaneous wire speed and its differential quantity, and the tension disturbance feature indirectly mapped by the speed fluctuation are extracted from the running speed data to constitute the original feature vector X4(t) of the fourth-class feature data; The wire bow feature vector is extracted from the bow height data to constitute the original feature vector X5(t) of the fifth-class feature data.

5. The multi-wire saw diamond wire active anti-breaking method according to claim 3, characterized in that: updating the target feature vector of each type of feature data The weights are in the following manner: The actual contribution of various types of feature data is defined as the feature contribution degree C i (t), wherein γ i is the feature variation sensitivity coefficient of the i-th feature data; The update weight formula is as follows: w i (t) = λ - w i (t - 1) + (1 - λ) - C i (t) Wherein, t-1 represents the last time; λ is a weight smoothing coefficient, and 0<λ<1; where the characteristic variation sensitivity coefficient γ i is calculated as follows: The absolute deviation sum of the real-time state vector and the reference state vector is calculated, and the calculation formula is as follows: wherein Z(t) is a real-time state vector, Z(t) = [Z1(t), Z2(t), Z3(t), Z4(t), Z5(t)], Z1(t) corresponds to the original feature vector X1(t) of the tension data, Z2(t) corresponds to the original feature vector X2(t) of the vibration data, Z3(t) corresponds to the original feature vector X3(t) of the temperature data, Z4(t) corresponds to the original feature vector X4(t) of the running speed data, and Z5(t) corresponds to the original feature vector X5(t) of the bow height data; Z ref = [Z 1,ref , Z 2,ref , ..., Z 5,ref ] is a reference vector under normal working conditions, and Z 1,ref is the mean value of the original feature vector X1(t), Z 2,ref is the mean value of the original feature vector X2(t), Z 3,ref is the mean value of the original feature vector X3(t), Z 4,ref is the mean value of the original feature vector X4(t), and Z 5,ref is the mean value of the original feature vector X5(t); ΔZ(t) is the Manhattan distance between Z(t) and Z ref , and ΔZ(t) is used to represent the absolute deviation sum of the real-time state vector and the reference state vector; The change amount of each type of feature data is calculated, and for the i-th type of feature data standardization feature, the change amount is Pearson's correlation coefficient is used to quantify the strength of the association between ΔZ(t): ρ i normalized with the initial value of the characteristic variation sensitivity coefficient γ i normalized with the initial value of the characteristic variation sensitivity coefficient γ i,init fusion, the final γi is obtained:

6. The multi-wire saw diamond wire active anti-breaking method according to claim 3, characterized in that: where the initial comprehensive feature vector F is obtained total in the manner of (t) as follows: Wherein θ(t) is a compensation term, and the calculation formula of θ(t) is as follows: where K is a correlation parameter, and 0 < K < 1 ; cov(X i ,X j ) is the covariance of the i-th feature data and the j-th feature data.

7. The multi-wire saw diamond wire active anti-breaking method according to claim 1, characterized in that: obtaining a target comprehensive feature vector in the following manner: Wherein N is the size of the sliding window; Based on target comprehensive feature vector The sequence of feature vectors X is computed as follows: The target integrated feature vector is calculated by continuously acquiring feature data with a time step of T at all discrete time points. Arranged chronologically, the final feature vector sequence X = [x1, x2, ..., x...] is obtained. t ,...,x T ]; Each element x in the feature vector sequence X t All are 5-dimensional feature vectors, and the overall sequence length is T.

8. The multi-wire saw diamond wire active anti-breaking method according to claim 1, characterized in that: The manner of combining LSTM and the attention mechanism to construct the hybrid deep learning model in step S3 is as follows: The feature vector sequence X is converted into the format required by the input layer of the LSTM in the prediction model; The LSTM performs time series feature extraction, and the LSTM network processes time series dependence through a forgetting gate, an input gate, and an output gate to generate a hidden state sequence H = [h1, h2,..., h t ,...,h T ] The attention mechanism acts on the sequence of hidden states H = [h1, h2,..., h t ,...,h T ] output by the LSTM, and highlights key time features by computing time step weights as follows: calculating an ability value e for measuring the importance of each hidden state t , and the calculation formula of the ability value e t is e t = a(h t , h T ) = W a · tanh(W h h t +W H h T +b a ) ; Then the e t softmax normalization is performed to obtain the attention weight α t , generating a context vector c, weighted fusion of hidden states wherein, T is a time step, h t is the output hidden state of the LSTM at the t time step; e t ∈R m is the energy value of the t time hidden state; a(·) is an attention score function, and an additive attention score function is adopted; α t ∈[0, 1] is the attention weight at the t time; c ∈R m is the context vector fused with the key features; Wa ∈R 1×m , W h ∈R m×n , W H ∈R m×n are weight matrices of the attention mechanism, respectively; b a ∈R m is a bias vector of the attention mechanism; The context vector c is input into the fully connected layer for feature conversion, and finally the prediction result R of the diamond wire breaking risk is obtained through the output layer.

9. The multi-wire saw diamond wire active anti-breaking method according to claim 1, characterized in that: The manner of adjusting the diamond wire based on the prediction result R in step S5 is as follows: calculating the maximum allowed tension T of the diamond wire max The calculation formula is as follows: where σ b is the tensile strength of the diamond wire, in N / m 2 ; d is the diameter of the diamond wire, in m; k s is the safety factor; calculating a target tension T of the diamond wire based on the prediction result R target The calculation formula is as follows: T target = T base · (1 - a · R) T min ≤T target ≤T max where T min is the relaxation threshold, in N; T base is the base tension, in N; and a is a risk adjustment factor, 0 < a < 1. The tension adjustment value ΔT is calculated based on the current tension and the target tension of the diamond wire, and the calculation formula of ΔT is as follows: ΔT = T target - T current where T current is the current tension, T current is monitored in real time by the tension sensor; The tension adjustment speed v is calculated based on the tension adjustment value ΔT T The tension adjustment speed v T The calculation formula is as follows: v T ≤v T,max wherein v T is the tension adjustment speed, in N / s; γ is the smoothing coefficient; t reg is the target adjustment time, in s; v T,max is the maximum adjustment speed, in N / s; Based on the tension adjustment value ΔT and the tension adjustment speed v T The tension adjustment value ΔT is corrected in real time, and the motor of the multi-wire cutting machine is controlled based on the PID control method. The calculation formula for controlling the motor operation is as follows: Wherein, u(t) is a motor control signal, used to drive the motor to adjust the tension; K p is a proportional gain of the PID control, i is an integral gain of the PID control, d is a differential gain of the PID control.

10. The active anti-breaking system of the diamond wire of the multi-wire saw according to any one of claims 1-9, characterized in that: The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device. The application relates to a multi-line cutting machine fracture risk prediction method and device.

Images