Broken line abnormity prediction method and device and computer equipment

Abstract

The invention provides a wire breakage abnormity prediction method and device and computer equipment, and particularly relates to the field of cutting wire cutting. The method comprises the following steps: acquiring cutting data of a previous cutting period; according to the main roller linear speed and the gradient of the main roller linear speed, segmenting a previous cutting period into a plurality of cutting working conditions, decomposing cutting process data into a plurality of sub-data sequences, and performing feature extraction on the sub-data sequence corresponding to each cutting working condition by adopting a convolutional neural network to obtain a time sequence feature of each cutting working condition; feature fusion is carried out on the time sequence features of the multiple cutting working conditions, and fused features are obtained; and obtaining a pre-trained reliability prediction model, and at least inputting the fused features and the abrasion loss of the cutting line into the reliability prediction model to obtain a reliability value of the cutting line in the next cutting period. According to the method, the cutting period is divided into a plurality of working conditions, and feature extraction and fusion are performed on the cutting data of the plurality of working conditions, so that the accuracy of broken line prediction is improved.

Description

Technical Field

[0001] The present application relates to the field of cutting wire cutting technology, and in particular to a method, device and computer equipment for predicting wire break anomaly. Background Art

[0002] In modern manufacturing, wire cutting technology is widely used in the processing of various materials as a precision processing method. This technology uses a thin cutting wire (such as diamond wire) moving at high speed and a cutting fluid to cut the workpiece. However, during the cutting process, the cutting wire is easily affected by various factors, such as workpiece installation accuracy, clamping force, cutting speed, tension fluctuations, and wear of the cutting wire itself, which may cause the cutting wire to break.

[0003] Once a wire break occurs, it will not only interrupt the processing process and reduce production efficiency, but may also cause damage to the workpiece and increase production costs. At present, most wire break monitoring methods in related technologies rely on single sensor monitoring or manual monitoring. These methods cannot effectively achieve real-time warning of wire break anomalies. Usually, they can only monitor and intervene after the wire break occurs, and cannot prevent the occurrence of wire break in advance. Summary of the invention

[0004] Based on this, it is necessary to provide a method, device and electronic device for predicting line break anomaly in response to the above technical problems, which can improve the accuracy and timeliness of line break anomaly prediction.

[0005] In a first aspect, the present application provides a method for predicting line disconnection anomalies, the method comprising: Acquire cutting data of the last cutting cycle, the cutting data at least including cutting process data and the wear amount of the cutting line, the cutting process data at least including the tension of the cutting line and the linear speed of the main roller; According to the main roller linear speed and the gradient of the main roller linear speed, the previous cutting cycle is divided into multiple cutting conditions, and according to the main roller linear speed and the gradient of the main roller linear speed of adjacent cutting conditions, the cutting process data is decomposed into multiple sub-data sequences, wherein each sub-data sequence corresponds to one cutting condition; A convolutional neural network is used to extract features from the sub-data sequence corresponding to each cutting condition to obtain the time series features of each cutting condition. Perform feature fusion on the time series features of multiple cutting conditions to obtain fused features; A pre-trained reliability prediction model is obtained, and at least the fused features and the wear amount of the cutting line are input into the reliability prediction model to obtain the reliability value of the cutting line for the next cutting cycle; wherein, the smaller the reliability value, the greater the possibility of the cutting line breaking abnormality.

[0006] In one embodiment, a convolutional neural network is used to extract features from the sub-data sequence corresponding to each cutting condition to obtain the time series features of each cutting condition, including: The sub-data sequence that is shorter than the set sequence length is padded through the padding operation; The parameters in the mask are used to mark whether the elements in the padded sub-data sequence are valid data. The parameters in the mask include a first parameter and a second parameter of different sizes. The first parameter is used to indicate that the element in the sub-data sequence corresponding to the position is valid data, and the second parameter indicates that the element in the sub-data sequence corresponding to the position is padding data. When using a convolutional neural network to extract features from the sub-data sequence corresponding to each cutting condition, the padded sub-data sequence is multiplied by the corresponding parameter in the mask to filter out the padding data so that the padded data does not participate in the convolution calculation.

[0007] In one embodiment, the time series features of multiple cutting conditions are fused to obtain fused features, including: For each cutting condition’s timing characteristics, the timing characteristics are mapped into a query matrix Q i , key matrix K i Sum Matrix V i , according to the mapped query matrix Q i , key matrix K i and the value matrix V i Get the new feature matrix for each cutting condition Z i ; According to the time order of each working condition in the previous cutting cycle, the new feature matrix is ​​connected end to end to obtain the fused feature; Among them, the query matrix Q i = W Q y i , the bond matrix K i = W K y i , value matrix V i = W V y i , W Q ,W K as well as W V are parameter matrices, y i Represents the timing characteristics corresponding to the i-th cutting condition.

[0008] In one embodiment, based on the mapped query matrix Q i , key matrix K i and the value matrix V i , the new feature matrix of each cutting condition is obtained by the following calculation method Z i : ; in, is the key matrix of the jth cutting condition. values, softmax is exponential normalization, the formula is: ; Among them, X represents The query matrix of the working condition and The product result sequence of the key matrix of the working condition, the product result sequence X is [ 1, 2,… l …, L ], L Indicates that the length of the product result sequence X is L .

[0009] In one embodiment, training a reliability prediction model includes: The reliability curve of the cutting line is constructed based on the cumulative distribution function of the Weibull distribution; Generate piecewise reliability function based on the reliability curve of the cutting line; Determine the label in the feature label pair based on the piecewise reliability function; the feature label pair includes a label and a feature, the feature in the feature label pair is the time series feature of each working condition in the historical cutting cycle, and the label in the feature label pair is the reliability value corresponding to the historical cutting cycle; The reliability prediction model is trained using feature label pairs, cutting line wear, and the actual position of the workbench; Among them, at least the fused features and the wear amount of the cutting line are input into the reliability prediction model to obtain the reliability value of the cutting line in the next cutting cycle, including: inputting the fused features, the wear amount of the cutting line and the current actual position of the workbench into the reliability prediction model to obtain the reliability value of the cutting line in the next cutting cycle.

[0010] In one embodiment, generating a piecewise reliability function based on the reliability curve of the cutting line includes: Determine a segmented reliability function according to the first time and the second time; wherein, if the use time of the cutting line is less than the first time, the reliability value of the cutting line is 1, and if the use time of the cutting line is greater than the second time, the reliability value of the cutting line is 0; Piecewise reliability function for: ; in, Indicates the first time, Indicates the second time, represents the usage time of the cutting line, k represents the shape parameter, and λ represents the scale parameter.

[0011] In one embodiment, the wear amount of the cutting wire is determined by the following formula: ; in, Indicates the amount of wear on the cutting line, idx Indicates the cycle number of the previous cutting cycle, cir_len Indicates the length of the cutting line around the main roller. len_n Indicates the net length of the cutting line during the last cutting cycle. v b Indicates the instantaneous speed of the worktable at the end of the previous cutting cycle. v l Indicates the instantaneous linear speed of the main roller at the end of the previous cutting cycle.

[0012] In one embodiment, the cutting working condition includes a wire feeding working condition and a wire taking-up working condition, the main roller linear speed of the wire feeding working condition is greater than zero, and the main roller linear speed of the wire taking-up working condition is less than zero; wherein the wire feeding working condition includes: Wire feeding acceleration condition: v>0 and f(v,t)>0; Wire feeding uniform cutting conditions: v>0 and f(v,t)=0; Wire feeding deceleration condition: v>0 and f(v,t)<0; The wire-receiving conditions include: Reeling acceleration condition: v<0 and f(v,t)<0; Rewinding uniform cutting conditions: v<0 and f(v,t)=0; Reeling deceleration condition: v<0 and f(v,t)>0; v represents the linear speed of the main roller, f(v,t) represents the gradient of the main roller linear velocity.

[0013] In a second aspect, the present application further provides a disconnection abnormality prediction device, the device comprising: An acquisition module, used for acquiring cutting data of the last cutting cycle, wherein the cutting data at least includes cutting process data and the wear amount of the cutting line, and the cutting process data at least includes the tension of the cutting line and the linear speed of the main roller; A segmentation module is used to segment the previous cutting cycle into multiple cutting conditions according to the main roller linear speed and the gradient of the main roller linear speed, and decompose the cutting process data into multiple sub-data sequences according to the main roller linear speed and the gradient of the main roller linear speed of adjacent cutting conditions, wherein each sub-data sequence corresponds to one cutting condition; A feature extraction module is used to extract features from the sub-data sequence corresponding to each cutting condition using a convolutional neural network to obtain the time series features of each cutting condition; A feature fusion module is used to fuse the time series features of multiple cutting conditions to obtain fused features; The prediction module is used to obtain a pre-trained reliability prediction model, at least input the fused features and the wear amount of the cutting line into the reliability prediction model, and obtain the reliability value of the cutting line in the next cutting cycle; wherein, the smaller the reliability value, the greater the possibility of the cutting line breaking abnormality.

[0014] In a third aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the disconnection anomaly prediction method of the first aspect is implemented.

[0015] The above-mentioned wire breakage abnormality prediction method obtains the cutting data of the previous cutting cycle, and the cutting data at least includes the tension of the cutting line, the linear speed of the main roller, and the wear of the cutting line. According to the changes in the linear speed of the main roller and its gradient, the previous cutting cycle is divided into multiple cutting conditions, and the cutting process data is decomposed into multiple sub-data sequences, each sub-sequence corresponding to a specific cutting condition. A convolutional neural network is used to extract features from the sub-data sequence of each cutting condition to obtain the time series features of each condition. The time series features are fused. The fused features and the wear of the cutting line are input into the pre-trained reliability prediction model to obtain the reliability value of the cutting line in the next cutting cycle. The smaller the reliability value, the greater the possibility of wire breakage abnormality in the cutting line. This method realizes the prediction of wire breakage abnormality by dividing the cutting cycle into multiple conditions, and extracts and fuses the cutting data of multiple conditions, thereby improving the accuracy of wire breakage prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flowchart of a disconnection abnormality prediction method in one embodiment; Figure 2 A schematic diagram of the linear speed and working condition division of the cutting process in one embodiment; Figure 3 It is a schematic diagram showing the trend of the output torque of the dual-side spindle motors in one embodiment; Figure 4 A flow chart of obtaining the timing characteristics of each cutting condition in one embodiment; Figure 5 A flow chart of fusing the time series features of multiple cutting conditions to obtain fused features in one embodiment; Figure 6 A flowchart of training a reliability prediction model in one embodiment; Figure 7 A reliability curve diagram in one embodiment; Figure 8 A diagram of a disconnection abnormality prediction device in an embodiment; Fig. 9 A structural diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] In one embodiment, Figure 1 As shown, a disconnection anomaly prediction method is provided, the method comprising the following steps: Step 101: obtaining cutting data of the last cutting cycle, wherein the cutting data at least includes cutting process data and the wear amount of the cutting line, and the cutting process data at least includes the tension of the cutting line and the linear speed of the main roller; It should be noted that a complete cutting operation includes thousands of cutting cycles, and the previous cutting cycle is one of the thousands of cutting cycles. The cutting data of the previous cutting cycle at least includes the cutting process data and the wear of the cutting wire, wherein the cutting wire can be a diamond wire, which is a wire with diamond particles fixed on the metal wire. The cutting process data mainly records the changes in various key parameters during the cutting process, including at least the tension of the cutting wire and the linear speed of the main roller. The cutting process data can also include the output torque of the double-sided spindle motor, the position of the worktable, the tension torque, etc.

[0019] It should be noted that the tension of the cutting line can refer to the tension on the left and right sides of the cutting line, that is, the pulling force on the left and right sides of the cutting line during the cutting process; the tension torque refers to the torque value used to control and maintain the tension of the cutting line during the cutting process; the output torque of the double-sided spindle motor refers to the torque output by the two-sided spindle motors when they are running in the cutting equipment. Furthermore, the main roller linear speed is the linear speed of the main roller driving the cutting line to perform reciprocating cutting motion, and its size and stability directly affect the cutting efficiency and quality. Among them, the main roller can be a roller wheel installed on the main shaft, which can rotate at high speed to drive the groove wheel and the wire mesh to rotate, contact with the silicon carbide crystal rod, and the crystal rod is cut into wafers through the grinding action of the cutting wire and the crystal rod.

[0020] Step 102: dividing the previous cutting cycle into a plurality of cutting conditions according to the main roller linear velocity and the gradient of the main roller linear velocity, and decomposing the cutting process data into a plurality of sub-data sequences according to the main roller linear velocity and the gradient of the main roller linear velocity of adjacent cutting conditions, wherein each sub-data sequence corresponds to a cutting condition; According to the changes in the main roller linear speed and its gradient, the working condition switching point in the cutting cycle can be determined. For example, when the main roller linear speed starts to increase from a positive value, it indicates that the wire feeding acceleration working condition has been entered; when the main roller linear speed reaches a stable value, it switches to the wire feeding uniform speed working condition; when the main roller linear speed begins to decrease, it enters the wire feeding deceleration working condition. When the main roller linear speed becomes a negative value and begins to increase the absolute value of the negative value, it enters the wire winding acceleration working condition; when the main roller linear speed stabilizes at a negative value, it is the wire winding uniform speed working condition; when the main roller linear speed decreases the absolute value of the negative value, it is the wire winding deceleration working condition.

[0021] According to the working condition switching point, the data sequence of the main roller linear speed of the cutting cycle can be divided into multiple sub-data sequences. Each sub-data sequence corresponds to a specific cutting working condition. For example, the sub-data sequence of the wire feeding acceleration working condition includes all data points from the beginning of the increase of the main roller linear speed to the main roller linear speed reaching a stable value; the sub-data sequence of the wire feeding uniform speed working condition includes all data points during the period when the main roller linear speed maintains a stable value; the sub-data sequence of the wire feeding deceleration working condition includes all data points from the beginning of the decrease of the main roller linear speed to the main roller linear speed dropping to zero. Similarly, the sub-data sequences of the wire winding acceleration working condition, the wire winding uniform speed working condition and the wire winding deceleration working condition are also divided in a similar manner.

[0022] Step 103: using a convolutional neural network to extract features from the sub-data sequence corresponding to each cutting condition, and obtaining the time series features of each cutting condition; The sub-data sequence of each cutting condition can be regarded as a time series data. The convolutional neural network can perform convolution operation on the sub-data sequence of each cutting condition through the convolution kernel in the convolution layer to extract features such as edges, trends, and periodicity. The size and number of the convolution kernels can be set according to the characteristics of the sub-data sequence.

[0023] In the convolution operation, the parameters (weights and biases) of the convolution kernel are learned and optimized through training, so that the convolutional neural network can automatically learn the most useful features for wire break prediction. After the convolution layer is processed, the sub-data sequence corresponding to each cutting condition is converted into the time series features of each cutting condition, which can more effectively reflect the dynamic changes in the cutting cycle.

[0024] It should be noted that in convolutional neural networks, the parameters of the convolution kernel mainly include weights and biases. Weights are numerical matrices used in the convolution kernel for element-by-element multiplication with the input data, and their size and shape are determined by the size of the convolution kernel. Bias is a scalar used to translate the result after the convolution operation to increase the flexibility of the model. The parameters of the convolution kernel can be learned and adjusted during the training process through optimization algorithms (such as gradient descent) so that the convolution kernel can effectively extract features from the sub-data sequence corresponding to each cutting condition.

[0025] Step 104: fusing the time series features of multiple cutting conditions to obtain fused features; There are various methods for feature fusion, the most common ones include feature concatenation, weighted averaging, and attention mechanism.

[0026] For example, suppose that the time series features of two cutting conditions are spliced, and the feature vectors of the wire feeding acceleration condition are F 1=[ f 11, f12 ,..., f 1n ] and the characteristic vector of the uniform wire feeding condition F 2=[ f 21 , f 22 ,..., f 2m ]. Perform feature fusion and fuse the two feature vectors sequentially into a longer feature vector, i.e. F =[ f 11 , f 12 ,..., f 1n , f 21 , f 22 ,..., f 2m ].

[0027] Step 105: Obtain a pre-trained reliability prediction model, input at least the fused features and the wear amount of the cutting line into the reliability prediction model, and obtain the reliability value of the cutting line for the next cutting cycle; wherein, the smaller the reliability value, the greater the possibility of the cutting line breaking abnormality.

[0028] The reliability prediction model is based on the internal deep learning structure, such as the long short-term memory network LSTM, etc., which performs complex nonlinear transformation and calculation on the input data and outputs a reliability value. The reliability value reflects the probability that the cutting wire can maintain normal operation in the next cutting cycle. The smaller the reliability value, the more obvious the abnormal characteristics of wire breakage during the cutting process, and the greater the possibility of wire breakage abnormality.

[0029] Exemplarily, an alarm threshold of the reliability value may be set in advance, for example, the alarm threshold is a reliability value of 0.1. If the reliability value output by the reliability prediction model is less than 0.1, a disconnection warning is issued.

[0030] In this embodiment, the method obtains the cutting data of the previous cutting cycle, and the cutting data includes at least the tension of the cutting line, the linear speed of the main roller, and the wear of the cutting line. According to the changes in the linear speed of the main roller and its gradient, the previous cutting cycle is divided into multiple cutting conditions, and the cutting process data is decomposed into multiple sub-data sequences, each sub-sequence corresponding to a specific cutting condition. A convolutional neural network is used to extract features from the sub-data sequence of each cutting condition to obtain the time series features of each condition. The time series features are fused. The fused features and the wear of the cutting line are input into the pre-trained reliability prediction model to obtain the reliability value of the cutting line in the next cutting cycle. The smaller the reliability value, the greater the possibility of the cutting line breaking abnormality. The method realizes the prediction of the broken wire abnormality and improves the accuracy of the broken wire prediction by dividing the cutting cycle into multiple conditions and extracting and fusing the cutting data of multiple conditions.

[0031] In one embodiment, the cutting working condition includes a wire feeding working condition and a wire taking-up working condition, the main roller linear speed of the wire feeding working condition is greater than zero, and the main roller linear speed of the wire taking-up working condition is less than zero; wherein the wire feeding working condition includes: wire feeding acceleration working condition: v>0 and f(v,t)>0; wire feeding uniform cutting condition: v>0 and f(v,t)=0; wire feeding deceleration condition: v>0 and f(v,t)<0; The wire-reeling conditions include: wire-reeling acceleration condition: v<0 and f(v,t)<0; Reeling-in uniform cutting condition: v<0 and f(v,t)=0; Reeling deceleration condition: v<0 and f(v,t)>0; Where, v represents the linear velocity of the main roller, f(v,t) represents the gradient of the main roller linear velocity.

[0032] like Figure 2 As shown, the main roller line speed can be v and its gradient The cutting process of a cutting cycle can be divided into multiple cutting conditions, where the horizontal axis is time t and the vertical axis is the main roller linear speed v. Each cutting condition has its own specific judgment logic. For example, the wire feeding acceleration condition corresponds to v >0 and f(v,t)>0, and the line reel deceleration condition corresponds to v <0 and f(v,t)>0.

[0033] Furthermore, by determining the segmentation nodes of different cutting conditions, the original continuous process data sequence can be segmented into several sub-data sequences according to the segmentation nodes. The trend of the cutting process data under different cutting conditions is as follows: Figure 3 It should be noted that Figure 3 Taking the output torque of the dual-side spindle motor as an example, different colors in the figure represent data points under different cutting conditions, where the horizontal axis is time and the vertical axis is the size of the data of the output torque of the dual-side spindle motor. Furthermore, under different cutting conditions, the trend and volatility characteristics of the cutting process data are very different, and the dynamics between different cutting conditions are strong. Therefore, for a complete cutting cycle X n,L , n is the number of cutting process data, L is the time span, and the cutting cycle is decomposed into 6 sub-data sequences according to the segmentation nodes of each working condition, which are X n,L1 , X n,L2 , X n,L3 , X n,L4 , X n,L5 , X n,L6 ,in L 1~ L 6 is the length of the sub-data sequence for each cutting condition.

[0034] In one embodiment, Figure 4 As shown, a convolutional neural network is used to extract features from the sub-data sequence corresponding to each cutting condition to obtain the time series features of each cutting condition, including the following steps: Step 401: fill in the sub-data sequence that is shorter than the set sequence length through a padding operation; It should be noted that each cutting cycle has a complete set of 6 cutting conditions, but the data lengths of the conditions with the same name in each cutting cycle are not exactly the same, and there are sampling errors. Length padding is required to make the data lengths of the conditions with the same name consistent.

[0035] In the process of cutting the cutting line, in order to eliminate the dimension and magnitude differences between different cutting process data, improve the comparability of data and the efficiency of model training, it is necessary to perform min-max normalization operation on each process data. Furthermore, the maximum number of elements in the sub-data sequence, i.e. tokens, is set. For sub-data sequences with data length less than the set value, a -1 value is filled at the end to ensure that all sub-data sequences input into the model have the same length.

[0036] For example, assuming that the longest element length of the sub-data sequence is set to 10, if a sub-data sequence has only 7 data points: [0.2, 0.5, 0.7, 0.9, 0.1, 0.3, 0.6], then 3 -1s are added to the end of the sub-data sequence to make its length reach 10, which becomes: [0.2, 0.5, 0.7, 0.9, 0.1, 0.3, 0.6, -1, -1, -1].

[0037] Step 402: Marking whether the element in the padded sub-data sequence is valid data by using the parameters in the mask, the parameters in the mask include a first parameter and a second parameter of different sizes, the first parameter is used to indicate that the element in the sub-data sequence corresponding to the position is valid data, and the second parameter indicates that the element in the sub-data sequence corresponding to the position is padding data; The mask is a sequence with the same length as the padded sequence, which contains two parameters: the first parameter and the second parameter. The first parameter can be 1, which indicates that the element at this position is valid data in the original sub-data sequence, and the second parameter can be 0, which indicates that the element at this position is padding data.

[0038] For example, assuming that the sub-data sequence becomes [0.2, 0.5, 0.7, 0.9, 0.1, 0.3, 0.6, -1, -1, -1] after normalization and padding, the corresponding mask should be [1, 1, 1, 1, 1, 1, 0, 0, 0]. In this mask, the 1s in the first 7 positions indicate that the corresponding data is valid data, and the 0s in the last 3 positions indicate that the corresponding data is -1, which is the padding data.

[0039] Step 403: When using the convolutional neural network to extract features from the sub-data sequence corresponding to each cutting condition, the padded sub-data sequence is multiplied by the corresponding parameter in the mask to filter out the padding data so that the padded data does not participate in the convolution calculation.

[0040] For example, assume that the padded sub-data sequence is: [0.2, 0.5, 0.7, 0.9, 0.1, 0.3, 0.6, -1, -1, -1], and the corresponding mask is: [1, 1, 1, 1, 1, 1, 0, 0, 0]. By multiplying the two sequences element by element, the resulting sequence is: [0.2, 0.5, 0.7, 0.9, 0.1, 0.3, 0.6, 0, 0, 0].

[0041] Furthermore, the padded -1 is multiplied by the 0 in the mask, and the resulting data is 0, that is, the padded data does not participate in the convolution kernel calculation. The valid data remains unchanged. During the convolution calculation process, the padded position corresponding to the 0 value will not have a substantial impact on the calculation of the convolution kernel, thereby avoiding the interference of the padded data on feature extraction.

[0042] In this embodiment, through padding operations and mask marking, sub-data sequences of different lengths can be effectively unified. By filtering out the padding data before convolution calculation, the accuracy and efficiency of feature extraction are guaranteed, and the interference of invalid data on model training and prediction is avoided.

[0043] In one embodiment, in order to effectively extract temporal features from sub-data sequences of different cutting conditions, six parallel feature extractors are constructed according to the type of cutting conditions, and each extractor consists of two temporal convolutional network (TCN) layers. Since the data characteristics under different cutting conditions are different, for example, the uniform cutting condition lasts for a long time and the timing feature fluctuation is small, and the timing feature fluctuation of the acceleration and deceleration condition is violent, different convolution kernel sizes can be set for the timing features. Under the uniform cutting condition, a larger convolution kernel size such as 5 is used to capture a wider range of time dependencies; while under the acceleration and deceleration cutting condition, a smaller convolution kernel size such as 3 is used to extract local detail information more finely. The temporal convolution kernel slides in a single step along the cutting time direction, and then extracts the timing features under 6 different cutting conditions in parallel. The feature extractor of each cutting condition can focus on its specific type of timing features, thereby improving the accuracy and efficiency of timing feature extraction.

[0044] In one embodiment, Figure 5 As shown, the time series features of multiple cutting conditions are fused to obtain fused features, including the following steps: Step 501: For each cutting condition's time series feature, map the time series feature into a query matrix Q i , key matrix K i Sum Matrix V i , according to the mapped query matrix Q i , key matrix K i and the value matrix V i Get the new feature matrix for each cutting condition Z i ; For the time series features extracted for the i-th cutting condition y 1, y 2, ..., y 6, each y i The shape is n × L i , n Indicates the number of cutting process data, L irepresents the length of the time series feature of the i-th cutting condition in the time dimension, that is, the number of time steps contained in the cutting condition. yi Mapped to query ( Q i = W Q y i ),key( K i = W K y i ) and value ( V i = W V y i ) three parts, among which W Q , W K , W V is the parameter matrix learned during training. W Q The matrix transforms the time series features into query representation. W K The matrix converts the time series features into key representations, W V The matrix converts the time series features into a value representation.

[0045] It should be noted that the time series characteristics of each cutting condition are mapped into three matrices, namely the query matrix Q i , key matrix K i Sum Matrix V i The mapping of these three matrices is achieved through a fully connected network, and the dimension and content of each matrix are related to the original time series features.

[0046] Query Matrix Q i Used to query the current time series feature and find the correlation with other time series features, which can help determine which other time series features are most relevant to the current time series feature. K i Used to match queries with other time series features to determine the correlation between time series features. Value matrix V i Contains the value information of the original time series features, which is used to generate new time series feature representations after correlation calculation.

[0047] By transforming the query matrix Qi and key matrix K i By performing matrix multiplication and applying the softmax function for normalization, we can get the attention score matrix, which represents the correlation weights between various temporal features. V i Multiply together to generate a new feature matrix for each cutting condition Z i The new feature matrix Z i The information of the original time series features is retained and the correlation between the fused time series features is obtained.

[0048] For example, assuming that the time series feature dimension of a cutting condition is n × m (in n is the number of features, m is the time step), after being mapped by the fully connected network, the query matrix Q i , key matrix K i Sum Matrix V i The dimensions are n × d , n × d and n × d (in d is the dimension after mapping). Through the attention calculation process, the new feature matrix obtained Z i The dimension is still n × m , the new feature matrix Z i It contains feature information adjusted by the attention mechanism, which can better reflect the importance and relationship of features in disconnection prediction.

[0049] Step 502: Connect the new feature matrices end to end according to the time order of the appearance of each working condition in the previous cutting cycle to obtain the fused features; wherein the query matrix Q i = W Q y i , the bond matrix K i = W K y i , value matrix V i = WV y i , W Q , W K as well as W V are parameter matrices, y i Represents the timing characteristics corresponding to the i-th cutting condition.

[0050] Each cutting condition corresponds to a new feature matrix Z i , these new feature matrices are obtained by processing the temporal features of each working condition through the attention mechanism. For example, assuming there are 6 working conditions, corresponding to 6 new feature matrices Z 1, Z 2, Z 3, Z 4, Z 5, Z 6. The order of the cutting conditions in the previous cutting cycle is wire feeding acceleration, wire feeding uniform speed, wire feeding deceleration, wire reeling acceleration, wire reeling uniform speed, and wire reeling deceleration. Connect the new feature matrix end to end according to this time sequence, that is, Z The end of 1 and Z 2 at the beginning of the connection, Z The end of 2 and Z 3, and so on, until all new feature matrices are connected in sequence to obtain a fused feature matrix Z n,L .

[0051] In this embodiment, the time series features under different cutting conditions are dynamically fused through the attention mechanism, which can more effectively capture the dynamic changes in the cutting process and the correlation between features. The fused features not only retain the information of the original time series features, but also highlight the feature parts that are more important for the wire break prediction task.

[0052] In one embodiment, according to the mapped query matrix Q i , key matrix K i and the value matrix V i , the new feature matrix of each cutting condition is obtained by the following calculation method Z i : ; in, is the key matrix of the jth cutting condition. values, softmax is exponential normalization, the formula is: ; Among them, X represents The query matrix of the working condition and The product result sequence of the key matrix of the working condition, the product result sequence X is [x1,x2,…x l …,x L ], x L Indicates that the length of the product result sequence X is L .

[0053] The query matrix of one of the cutting conditions Q i Bond matrix with all other cutting conditions K j No. l The values ​​are multiplied, where l Represents the key matrix K j The l The purpose of the product operation is to measure the query matrix Q i With key matrix K j The relevance at a specific location is measured in the query matrix Q i With the measure key matrix K j The l Further, the query matrix Q i Bond matrix with all other cutting conditions K j No. l All product results obtained by multiplying values ​​can form a product result sequence X .

[0054] Furthermore, for the product result sequence X Use the softmax function for normalization. The softmax function can convert the product result sequence X Each value in is converted into a weight value between 0 and 1. Further, each weight value after normalization obtained by using the softmax function is compared with the corresponding value matrix V j Perform weighted summation to obtain the new feature matrix corresponding to the cutting condition Z i .

[0055] It should be noted that exponential normalization is performed on each value in the product result sequence X, which means that all exponential terms are added together to obtain the denominator , ensuring that the output of the softmax function is a probability distribution. l The weight of the position Calculated as: . Indicates that in the product result sequence X l The relative importance of each position. The larger the weight value, the more important the position.

[0056] For example, assuming there are 6 cutting conditions, the query matrix Q for each condition is i , key matrix K i and the value matrix V i All of them have been mapped through the fully connected network. For the first cutting condition, its query matrix Q1 is respectively connected to the key matrices K1 to K6 of all six conditions. l The 6 values ​​are multiplied to obtain 6 values, which can form the product result sequence X. The 6 values ​​in the product result sequence X are normalized using the softmax function to obtain 6 weight values. Among them, the 6 weight values ​​are between 0 and 1. The 6 weight values ​​are weighted and summed with the value matrices V1 to V6 under the corresponding cutting conditions to obtain the new feature matrix Z1 of the first cutting condition. The same process is also applied to the other 5 cutting conditions, and finally the new feature matrices Z1 to Z6 of all 6 cutting conditions are obtained.

[0057] In this embodiment, the new feature matrix of each cutting condition can dynamically integrate the feature information of other conditions and adjust the weights according to the correlation in the wire break prediction task, thereby improving the richness of feature representation and the accuracy of the prediction model.

[0058] In one embodiment, Figure 6 As shown, training the reliability prediction model includes the following steps: Step 601: constructing a reliability curve of the cutting line based on the cumulative distribution function of the Weibull distribution; The reliability curve constructed by Weibull distribution is as follows Figure 7 As shown, the horizontal axis is the cutting time in seconds (s), and the vertical axis is the reliability value. Weibull distribution is a commonly used life distribution model. The life degradation curve of the cutting wire is based on the cumulative distribution function (CDF) of Weibull distribution. The cutting data of wire breakage in historical records is used as the research object. The reliability of the normal cutting period is set to 1. As the cutting wire wears and degrades, the life gradually declines until the reliability is set to 0 when the wire breakage occurs.

[0059] When constructing the reliability curve of the cutting line, the cumulative distribution function based on the Weibull distribution can effectively characterize the reliability change trend of the cutting line during use. The cumulative distribution function (CDF) expression of the Weibull distribution is: , in, x is a non-negative random variable, usually representing time or life span data; k is a shape parameter. For wire breakage caused by cutting wire wear, the equipment failure rate and failure rate gradually increase over time, and the value is generally 1.5; λ is a scale parameter, which affects the scale of the distribution and represents the characteristic life of the product.

[0060] It should be noted that by recording the time T of each wire breakage e And the time T when the process parameters begin to become abnormal s , the time difference Δt=T is calculated e -T s Δt can reflect the time span from the occurrence of anomaly to the final wire break. After further collecting Δt data of multiple cutting cycles, the maximum likelihood estimation method can be used to fit the Δt data of multiple cutting cycles into a Weibull distribution, thereby solving the scale parameter λ.

[0061] Step 602: Generate a piecewise reliability function based on the reliability curve of the cutting line; Collect data from the actual use of the cutting wire, including normal working time and the time when the wire breaks. Perform statistical analysis on these data and use the life distribution model such as Weibull distribution to determine the shape parameter k and scale parameter λ of each stage. Furthermore, through the shape parameter k and scale parameter λ, a segmented reliability function can be constructed. The function form of each stage is R(t)= , where t is time, λ is the scale parameter, and k is the shape parameter. The piecewise function can more accurately describe the reliability change trend of the cutting line at different stages.

[0062] Step 603: determining the label in the feature label pair based on the piecewise reliability function; the feature label pair includes a label and a feature, the feature in the feature label pair is the time series feature of each working condition in the historical cutting cycle, and the label in the feature label pair is the reliability value corresponding to the historical cutting cycle; The time series feature can reflect the dynamic changes of the cutting process and the wear of the cutting line. The label can be determined based on the piecewise reliability function, which is generated based on the reliability data of the cutting line at different stages and can more accurately describe the reliability changes of the cutting line under different usage times or conditions. Substituting the time data of each historical cutting cycle into the corresponding piecewise reliability function, the corresponding reliability value can be obtained as the label in the feature label pair.

[0063] Step 604: Use feature label pairs, cutting line wear and the actual position of the workbench to train a reliability prediction model; wherein at least the fused features and the cutting line wear are input into the reliability prediction model to obtain the reliability value of the cutting line for the next cutting cycle, including: inputting the fused features, the cutting line wear and the current actual position of the workbench into the reliability prediction model to obtain the reliability value of the cutting line for the next cutting cycle.

[0064] The feature label pair may include the timing characteristics of each cutting condition in the historical cutting cycle and the corresponding reliability value. By learning the relationship between the timing characteristics and reliability values ​​of each cutting condition in the historical cutting cycle, the reliability prediction model can understand the patterns that affect the reliability of the cutting line under different cutting conditions.

[0065] The amount of wear on the cutting wire can directly reflect the degree of loss of the cutting wire during the cutting process. As the cutting progresses, the cutting wire will gradually wear, its diameter will decrease, and its strength will decrease, thereby increasing the risk of wire breakage. By taking the amount of wear on the cutting wire as the input of the reliability prediction model, the reliability prediction model can more accurately evaluate the remaining life and reliability of the cutting wire. The calculation of the amount of wear is based on the length of use of the cutting wire in each cutting cycle, the length of one circle of the main roller, the worktable speed, and the main roller linear speed, etc. These parameters can accurately quantify the loss of the cutting wire.

[0066] The workbench is a component that fixes the crystal rod and feeds it slowly downward. During the cutting process, the stability of the current actual position of the workbench has a great influence on the cutting effect. The actual position of the workbench can be monitored in real time by a sensor or encoder, reflecting the precise coordinates or position information of the workbench during the cutting process. When the workbench is fed to the position where the contact length between the crystal rod and the wire mesh is the longest, the cutting resistance is the greatest. At this time, if the workbench is vibrating and unstable, wire breakage is likely to occur.

[0067] In the reliability prediction model training phase, the fused features, the wear of the cutting line, and the current actual position of the workbench are used as input. By selecting a suitable algorithm, such as the long short-term memory network (LSTM), and using the time series features and reliability values ​​of each cutting condition in the historical cutting cycle for training, the reliability prediction model can learn the relationship between the time series features and the reliability value of the cutting line. During the training process, the sliding window is continuously slid along the cutting time direction to construct a mapping combination of feature labels, and the parameters of the model are continuously adjusted to minimize the error between the predicted value and the actual value until the reliability prediction model shows good prediction performance.

[0068] The trained reliability prediction model can predict the reliability of the cutting line in the next cutting cycle based on the timing characteristics of different cutting conditions in the current cutting cycle during the actual cutting process.

[0069] In one embodiment, based on the reliability curve of the cutting line, a piecewise reliability function is generated, including: Determine a segmented reliability function according to the first time and the second time; wherein, if the use time of the cutting line is less than the first time, the reliability value of the cutting line is 1, and if the use time of the cutting line is greater than the second time, the reliability value of the cutting line is 0; Piecewise reliability function for: ; in, Indicates the first time, Indicates the second time, represents the usage time of the cutting line, k represents the shape parameter, and λ represents the scale parameter.

[0070] Assume there is a set of data about the failure time of the cutting line, recorded as t 1, t 2,..., t n , the value of λ can be estimated using the following formula: λ = .

[0071] in, k is the shape parameter and can be estimated by maximum likelihood estimation or other methods.

[0072] Determine two time points, namely the first time T s and the second time T e When the cutting line is used for a long time t No more than the first time T s When the usage time is , it can be considered that the cutting line is in the initial use stage and the reliability is maintained at the highest level, so its reliability value is set to 1; t Exceed T e When , the cutting wire has entered the excessive wear stage, and its reliability is significantly reduced. At this time, its reliability value is set to 0. T s and T e In the stage between , the reliability of the cutting line shows a gradual downward trend. The reliability value of this stage can be calculated by the cumulative distribution function of the Weibull distribution, and the formula is: R ( t )= ,in k is a shape parameter, reflecting the trend of failure rate over time. λis the scale parameter, which represents the characteristic life. The entire piecewise reliability function R It can be expressed as: ; In practical applications, T s and T e The determination of the actual service life of the cutting wire needs to be analyzed based on a large amount of cutting data. By collecting the time data from the beginning of use of the cutting wire to the occurrence of wire breakage in multiple cutting cycles, as well as the time data when the process parameters begin to become abnormal, the statistical analysis method can be used to determine the actual service life of the cutting wire. T s and T e Further, the maximum likelihood estimation method is used to fit the Weibull distribution based on these data to obtain the shape parameter k and scale parameter λ The best estimate of .

[0073] In one embodiment, the wear amount of the cutting wire is determined by the following formula; ; in, Indicates the amount of wear on the cutting line, idx Indicates the cycle number of the previous cutting cycle, cir_len Indicates the length of the cutting line around the main roller. len_n Indicates the net length of the cutting line during the last cutting cycle. v b Indicates the instantaneous speed of the worktable at the end of the previous cutting cycle. v l Indicates the instantaneous speed of the main roller line at the end of the previous cutting cycle.

[0074] wear Indicates the wear of the cutting line, which can be obtained by accumulating the wear of each cutting cycle. idx Indicates the cycle number of the last cutting cycle, which is used to determine the range of accumulated wear. cir _ len The length of the cutting line around the main roller can be expressed, and the wear of the cutting line in each cycle can be calculated. lenn It can indicate the net length of the cutting line in the last cutting cycle, and can reflect the actual length of the cutting line used in the cycle. V b It can indicate the instantaneous speed of the worktable at the end of the previous cutting cycle. v lIt can represent the instantaneous speed of the main roller line at the end of the previous cutting cycle. These two speed parameters can jointly affect the wear rate of the cutting wire during the cutting process. The workbench moves slowly downward during the cutting process, thereby pressing the crystal rod against the wire mesh and giving the crystal rod a downward feed force. Therefore, the workbench has an instantaneous speed during the cutting process.

[0075] Furthermore, by substituting these parameters into the formula and accumulating them for all cutting cycles, the total wear of the cutting wire can be obtained: wear , thereby providing important data support for subsequent reliability prediction and line disconnection warning.

[0076] In this embodiment, this division method helps to more accurately analyze and control different stages of the cutting process.

[0077] Based on the same concept, Figure 8 As shown, the present application also provides a disconnection abnormality prediction device, the device comprising: An acquisition module 801 is used to acquire cutting data of the last cutting cycle, wherein the cutting data at least includes cutting process data and the wear amount of the cutting line, and the cutting process data at least includes the tension of the cutting line and the linear speed of the main roller; The segmentation module 802 is used to segment the previous cutting cycle into multiple cutting conditions according to the main roller linear speed and the gradient of the main roller linear speed, and decompose the cutting process data into multiple sub-data sequences according to the main roller linear speed and the gradient of the main roller linear speed of adjacent cutting conditions, wherein each sub-data sequence corresponds to one cutting condition; The feature extraction module 803 is used to extract features from the sub-data sequence corresponding to each cutting condition using a convolutional neural network to obtain the time series features of each cutting condition; A feature fusion module 804 is used to fuse the time series features of multiple cutting conditions to obtain fused features; The prediction module 805 is used to obtain a pre-trained reliability prediction model, at least input the fused features and the wear amount of the cutting line into the reliability prediction model, and obtain the reliability value of the cutting line in the next cutting cycle; wherein, the smaller the reliability value, the greater the possibility of the cutting line breaking abnormality.

[0078] In one embodiment, the feature extraction module 803 uses a convolutional neural network to perform feature extraction on the sub-data sequence corresponding to each cutting condition to obtain the timing features of each cutting condition, which is specifically used for: padding the sub-data sequence that is less than the set sequence length through a padding operation; marking the elements in the padded sub-data sequence through the parameters in the mask to determine whether they are valid data, the parameters in the mask include a first parameter and a second parameter of different sizes, the first parameter is used to indicate that the element in the sub-data sequence corresponding to the position is valid data, and the second parameter indicates that the element in the sub-data sequence corresponding to the position is padding data; when using a convolutional neural network to perform feature extraction on the sub-data sequence corresponding to each cutting condition, the padding data is filtered out by multiplying the padded sub-data sequence with the corresponding parameter in the mask so that the padded data does not participate in the convolution calculation.

[0079] In one embodiment, the feature fusion module 804 performs feature fusion on the time series features of multiple cutting conditions to obtain fused features, which are specifically used to: for each time series feature of the cutting condition, map the time series feature into a query matrix Q i , key matrix K i Sum Matrix V i , according to the mapped query matrix Q i , key matrix K i and the value matrix V i Get the new feature matrix for each cutting condition Z i ; According to the time order of each working condition in the previous cutting cycle, the new feature matrix is ​​connected end to end to obtain the fused feature; among them, the query matrix Q i = W Q y i , the bond matrix K i = W K y i , value matrix V i = W V y i , W Q , W K as well as W V are parameter matrices,y i Represents the timing characteristics corresponding to the i-th cutting condition.

[0080] In one embodiment, the feature fusion module 804, based on the mapped query matrix Q i , key matrix K i and the value matrix V i , the new feature matrix of each cutting condition is obtained by the following calculation method Z i : ;in, is the key matrix of the jth cutting condition. values, softmax is exponential normalization, the formula is: ; Where x represents the product result sequence of the query matrix of one cutting condition and the key matrix of the condition. The product result sequence x is [x1, x2, …x l …,x L ], x L Indicates that the length of the product result sequence x is L .

[0081] In one embodiment, the prediction module 805 trains a reliability prediction model, which is specifically used to: construct a reliability curve of the cutting line based on the cumulative distribution function of the Weibull distribution; generate a piecewise reliability function based on the reliability curve of the cutting line; determine the label in the feature label pair based on the piecewise reliability function; the feature label pair includes a label and a feature, the feature in the feature label pair is the time series feature of each working condition in the historical cutting cycle, and the label in the feature label pair is the reliability value corresponding to the historical cutting cycle; use the feature label pair, the wear amount of the cutting line and the actual position of the workbench to train the reliability prediction model; wherein, at least the fused feature and the wear amount of the cutting line are input into the reliability prediction model to obtain the reliability value of the cutting line in the next cutting cycle, including: inputting the fused feature, the wear amount of the cutting line and the current actual position of the workbench into the reliability prediction model to obtain the reliability value of the cutting line in the next cutting cycle.

[0082] In one embodiment, the prediction module 805 generates a piecewise reliability function based on the reliability curve of the cutting line, which is specifically used to: determine the piecewise reliability function according to the first time and the second time; wherein, if the use time of the cutting line is less than the first time, the reliability value of the cutting line is 1, and if the use time of the cutting line is greater than the second time, the reliability value of the cutting line is 0; the piecewise reliability function for: ; in, Indicates the first time, Indicates the second time, represents the usage time of the cutting line, k represents the shape parameter, and λ represents the scale parameter.

[0083] In one embodiment, the prediction module 805 determines the wear amount of the cutting line by the following formula: ; in, Indicates the amount of wear on the cutting line, idx Indicates the cycle number of the previous cutting cycle, cir_len Indicates the length of the cutting line around the main roller. len_n Indicates the net length of the cutting line during the last cutting cycle. v b Indicates the instantaneous speed of the worktable at the end of the previous cutting cycle. v l Indicates the instantaneous speed of the main roller line at the end of the previous cutting cycle.

[0084] In one embodiment, the cutting conditions of the segmentation module 802 include a wire feeding condition and a wire taking-up condition, wherein the main roller linear velocity of the wire feeding condition is greater than zero, and the main roller linear velocity of the wire taking-up condition is less than zero; wherein the wire feeding condition includes: wire feeding acceleration condition: v>0 and f(v,t)>0; wire feeding uniform cutting condition: v>0 and f(v,t)=0; wire feeding deceleration condition: v>0 and f(v,t)<0; The wire-reeling working conditions include: wire-reeling acceleration working conditions: v<0 and f(v,t)<0; Reeling-in uniform cutting condition: v<0 and f(v,t)=0; Reeling deceleration condition: v<0 and f(v,t)>0; v represents the linear velocity of the main roller, f(v,t) represents the gradient of the main roller linear velocity.

[0085] Based on the same concept, the present application also provides a computer device, including a memory and a processor. The computer device may be a terminal, and its internal structure diagram may be as shown in FIG. Fig. 9As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. When the computer program is executed by the processor, the disconnection anomaly prediction method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0086] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0087] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0088] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A disconnection anomaly prediction method, characterized in that: The method comprises: Acquire cutting data of the last cutting cycle, wherein the cutting data at least includes cutting process data and wear of the cutting line, wherein the cutting process data at least includes the tension of the cutting line and the linear speed of the main roller; According to the main roller linear speed and the gradient of the main roller linear speed, the previous cutting cycle is divided into a plurality of cutting conditions, and according to the main roller linear speed and the gradient of the main roller linear speed of adjacent cutting conditions, the cutting process data is decomposed into a plurality of sub-data sequences, wherein each sub-data sequence corresponds to one cutting condition; A convolutional neural network is used to extract features from the sub-data sequence corresponding to each cutting condition to obtain the time series features of each cutting condition; Performing feature fusion on the time series features of the plurality of cutting conditions to obtain fused features; A pre-trained reliability prediction model is obtained, and at least the fused features and the wear amount of the cutting line are input into the reliability prediction model to obtain a reliability value of the cutting line for the next cutting cycle; wherein, the smaller the reliability value, the greater the possibility that the cutting line will break abnormally.

2. The disconnection abnormality prediction method according to claim 1, characterized in that: A convolutional neural network is used to extract features from the sub-data sequence corresponding to each cutting condition to obtain the time series features of each cutting condition, including: The sub-data sequence that is shorter than the set sequence length is padded through the padding operation; Marking whether the elements in the padded sub-data sequence are valid data by using the parameters in the mask, wherein the parameters in the mask include a first parameter and a second parameter of different sizes, wherein the first parameter is used to indicate that the element in the sub-data sequence corresponding to the position is valid data, and the second parameter indicates that the element in the sub-data sequence corresponding to the position is padding data; When a convolutional neural network is used to extract features from a sub-data sequence corresponding to each cutting condition, the padded sub-data sequence is multiplied by the corresponding parameter in the mask to filter out the padding data so that the padded data does not participate in the convolution calculation.

3. The disconnection abnormality prediction method according to claim 1, characterized in that: The time series features of the multiple cutting conditions are fused. The fused features include: For each cutting condition, the timing characteristics are mapped into a query matrix Q i , key matrix K i Sum Matrix V i , according to the query matrix mapped Q i , key matrix K i and the value matrix V i Get the new feature matrix for each cutting condition Z i ; According to the time order in which each working condition appears in the previous cutting cycle, the new feature matrix is ​​connected end to end to obtain a fused feature; Among them, the query matrix Q i = W Q y i , the bond matrix K i = W K y i , the value matrix V i = W V y i , W Q , W K as well as W V are all parameter matrices. y i Represents the timing characteristics corresponding to the i-th cutting condition.

4. The disconnection abnormality prediction method according to claim 3, characterized in that: According to the query matrix mapped Q i , key matrix K i and the value matrix V i , the new feature matrix of each cutting condition is obtained by the following calculation method Z i : ; in, is the key matrix of the jth cutting condition. l values, softmax is exponential normalization, the formula is: ； Among them, X represents The query matrix of the working condition and The product result sequence of the key matrix of the working condition, the product result sequence X is [ 1, 2,… l …, L ], L Indicates that the length of the product result sequence X is L .

5. The disconnection abnormality prediction method according to claim 1, characterized in that: Training the reliability prediction model includes: Constructing a reliability curve of the cutting line based on a cumulative distribution function of Weibull distribution; Based on the reliability curve of the cutting line, generating a piecewise reliability function; Determine the label in the feature label pair based on the segmented reliability function; the feature label pair includes a label and a feature, the feature in the feature label pair is the time series feature of each working condition in the historical cutting cycle, and the label in the feature label pair is the reliability value corresponding to the historical cutting cycle; The reliability prediction model is trained by using the feature label pair, the wear amount of the cutting line and the actual position of the workbench; Wherein, at least the fused features and the wear amount of the cutting line are input into the reliability prediction model to obtain the reliability value of the cutting line for the next cutting cycle, including: inputting the fused features, the wear amount of the cutting line and the current actual position of the workbench into the reliability prediction model to obtain the reliability value of the cutting line for the next cutting cycle.

6. The disconnection abnormality prediction method according to claim 5, characterized in that: Based on the reliability curve of the cutting line, a piecewise reliability function is generated, including: Determine a piecewise reliability function according to the first time and the second time; wherein, if the use time of the cutting line is less than the first time, the reliability value of the cutting line is 1, and if the use time of the cutting line is greater than the second time, the reliability value of the cutting line is 0; The piecewise reliability function for: ; in, Indicates the first time, Indicates the second time, represents the usage time of the cutting line, k represents the shape parameter, and λ represents the scale parameter.

7. The disconnection abnormality prediction method according to claim 1, characterized in that: The wear amount of the cutting line is determined by the following formula; ； in, Indicates the amount of wear on the cutting line, idx represents the cycle number of the previous cutting cycle, cir_len Indicates the length of the cutting line around the main roller. len_n Indicates the net length of the cutting line during the previous cutting cycle. v b Indicates the instantaneous speed of the worktable at the end of the previous cutting cycle, v l It represents the instantaneous linear speed of the main roller at the end of the previous cutting cycle.

8. The method for predicting line break anomaly according to any one of claims 1 to 7, characterized in that: The cutting working condition includes a wire feeding working condition and a wire taking-up working condition, wherein the main roller linear speed of the wire feeding working condition is greater than zero, and the main roller linear speed of the wire taking-up working condition is less than zero; wherein the wire feeding working condition includes: Wire feeding acceleration condition: v>0 and f(v,t)>0; Wire feeding uniform cutting conditions: v>0 and f(v,t)=0; Wire feeding deceleration condition: v>0 and f(v,t)<0; The wire-receiving working conditions include: Reeling acceleration condition: v<0 and f(v,t)<0; Rewinding uniform cutting conditions: v<0 and f(v,t)=0; Reeling deceleration condition: v<0 and f(v,t)>0; v represents the linear speed of the main roller, f(v,t) represents the gradient of the main roller linear velocity.

9. A disconnection abnormality prediction device, characterized in that: The device comprises: An acquisition module, used for acquiring cutting data of the last cutting cycle, wherein the cutting data at least includes cutting process data and wear of the cutting line, wherein the cutting process data at least includes the tension of the cutting line and the linear speed of the main roller; a segmentation module, configured to segment the previous cutting cycle into a plurality of cutting conditions according to the main roller linear speed and the gradient of the main roller linear speed, and decompose the cutting process data into a plurality of sub-data sequences according to the main roller linear speed and the gradient of the main roller linear speed of adjacent cutting conditions, wherein each sub-data sequence corresponds to one cutting condition; A feature extraction module, used for extracting features from the sub-data sequence corresponding to each cutting condition using a convolutional neural network to obtain a time series feature of each cutting condition; A feature fusion module, used for fusing the time series features of the plurality of cutting conditions to obtain fused features; A prediction module is used to obtain a pre-trained reliability prediction model, input at least the fused features and the wear amount of the cutting line into the reliability prediction model, and obtain the reliability value of the cutting line in the next cutting cycle; wherein, the smaller the reliability value, the greater the possibility of the cutting line breaking abnormality.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the disconnection abnormality prediction method according to any one of claims 1 to 8 are implemented.

Images