Intelligent production management method and system for p-type product melt spot welding and medium

Abstract

The application provides a kind of intelligent production management method, system and medium for P-type product melt spot welding, comprising: collecting spot welding current trigger time, electrode pressure, nugget formation time domain signal, extracting feature vector by means of double time domain window, calculating phase difference value ratio to obtain adjustable second threshold;Feature vector is stored in asynchronous cache queue, priority index is set by combining parameter deviation and phase difference value ratio, and original timing is maintained;Monitor queue memory occupancy rate, trace of sliding cumulative deviation matrix, and start synchronous strong persistence when exceeding threshold value;When frame loss occurs, build compensation matrix by means of adjacent effective frame phase offset, interpolate to complete missing frame timestamp, then adjust signal interception window length according to matrix main eigenvalue, complete data archiving.

Description

Technical Field

[0001] This application belongs to the field of melt spot welding, and in particular relates to an intelligent production management method, system and medium for melt spot welding of P-type products. Background Technology

[0002] In the semiconductor and precision electronic component manufacturing industry, P-type products, due to their unique properties, have high requirements for the fusion spot welding process. During spot welding, there is a fusion relationship between current pulses, electrode pressure, and the formation state of the weld nugget. The acquisition and management of this production data is a core link in ensuring consistent welding quality and enabling process backtracking. Existing production data management methods typically use traditional data acquisition modules to acquire electrical and mechanical signals during welding, extract features by capturing data through a preset sliding window, and store the data in a database using conventional asynchronous or synchronous writing methods. To alleviate the storage pressure caused by instantaneous high-frequency data, some systems utilize asynchronous caching mechanisms, which temporarily store the acquired data in a memory queue and then perform batch persistence operations to complete archiving when system resources are idle or predetermined conditions are met. However, existing asynchronous caching mechanisms struggle to identify the urgency of high-deviation abnormal data when spot welding P-type products, and also fail to incorporate the phase relationship and cumulative deviation of multi-source signals into the persistence strategy. This leads to the inability to identify and record key information in a timely manner during critical process fluctuations, easily resulting in the omission of important features. Furthermore, it is difficult to adaptively adjust the subsequent data extraction strategy based on the degree of data frame loss, resulting in disordered timestamps in the reconstructed data and distorted subsequent archived data, affecting the integrity of overall production data management and the accuracy of process analysis. Summary of the Invention

[0003] According to one aspect of the present invention, an intelligent production management method for melt spot welding of P-type products is proposed to solve the problem that existing methods cannot adaptively adjust subsequent data retrieval strategies according to the degree of data loss, resulting in disordered timestamps in the reconstructed data and distortion of subsequent archived data. The method includes the following steps: Production data of the spot welding process is collected, including the spot welding current pulse trigger time, electrode pressure time series and weld nugget formation time domain signal. Based on the trigger time, the time series and the time domain signal are subjected to dual time domain window truncation to extract feature vectors. The phase difference ratio is calculated according to the time series and the time domain signal, and the adjusted second threshold is obtained based on the phase difference ratio. The feature vector is pushed into an asynchronous cache queue. A weight is calculated based on the absolute value of the deviation between the feature vector and the standard melt parameters and the ratio of the phase difference. A priority read index is constructed for the queue based on the weight without changing the queue's temporal order. The memory occupancy rate of the queue is monitored. A sliding cumulative deviation matrix is ​​calculated based on the feature vectors that maintain the temporal order. When the occupancy rate is greater than a preset first threshold or the trace of the matrix is ​​greater than the adjusted second threshold, the system switches to synchronous strong persistence mode. In the synchronous strong persistence mode, the read order of the feature vectors to be persisted is determined according to the priority read index. When it is determined that a data frame is lost, a compensation matrix is ​​constructed based on the electrode pressure phase offset of the adjacent valid frames. The timestamps of the lost data frames are interpolated and re-indexed according to the compensation matrix to generate a dataset. The window truncation length of subsequent production data is scaled based on the principal eigenvalue of the compensation matrix and then archived to the database.

[0004] According to another aspect of the present invention, an intelligent production management system for melt spot welding of P-type products is proposed, comprising the following modules: The acquisition module is used to acquire production data of the spot welding process, including the spot welding current pulse trigger time, electrode pressure time series and weld nugget formation time domain signal. Based on the trigger time, the time series and the time domain signal are subjected to dual time domain window truncation to extract feature vectors. The phase difference ratio is calculated according to the time series and the time domain signal, and the adjusted second threshold is obtained based on the phase difference ratio. The module is used to push the feature vector into an asynchronous cache queue, calculate the weight based on the absolute value of the deviation between the feature vector and the standard melt parameters and the ratio of the phase difference, build a priority read index for the queue based on the weight without changing the queue's temporal order, monitor the memory usage of the queue, calculate the sliding cumulative deviation matrix based on the feature vector while maintaining the temporal order, and switch to synchronous strong persistence mode when the memory usage exceeds a preset first threshold or the trace of the matrix exceeds the adjusted second threshold. In the synchronous strong persistence mode, the reading order of the feature vectors to be persisted is determined according to the priority read index. The generation module is used to construct a compensation matrix based on the electrode pressure phase offset of adjacent valid frames when it is determined that a data frame is lost. It then interpolates and re-indexes the timestamps of the lost data frames according to the compensation matrix to generate a dataset. Finally, it scales the window length of subsequent production data based on the principal eigenvalues ​​of the compensation matrix and archives it into the database.

[0005] This invention improves data processing accuracy by acquiring multi-source signals from the spot welding process and performing dual-time-domain windowing, combined with calculating the phase difference ratio to extract feature vectors and setting targeted judgment thresholds. A weighted priority index based on parameter deviation and phase difference ratio is introduced into the asynchronous cache queue. While maintaining the original data sequence, it switches to a synchronous strong persistence mode at critical nodes by monitoring memory usage and the trace of the sliding cumulative deviation matrix, balancing the throughput efficiency of high-concurrency data with the security of core data storage. Furthermore, for data frame loss, a compensation matrix of adjacent valid frames is constructed for timestamp interpolation and re-indexing, and the truncation length of subsequent data is scaled according to the principal feature value, ensuring the integrity and coherence of the archived dataset and improving the quality monitoring capabilities of spot welding production management. Attached Figure Description

[0006] Figure 1 A flowchart of the first embodiment; Figure 2 This is a heatmap of the phase offset compensation matrix. Detailed Implementation

[0007] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0008] In the first embodiment, the present invention proposes an intelligent production management method for melt spot welding of P-type products, such as... Figure 1 It includes the following steps: S1. Collect production data of the spot welding process, including the spot welding current pulse trigger time, electrode pressure time series and weld nugget formation time domain signal. Based on the trigger time, perform dual time domain window extraction on the time series and the time domain signal to extract feature vectors. Calculate the phase difference ratio based on the time series and the time domain signal, and obtain the adjusted second threshold based on the phase difference ratio.

[0009] A high-frequency Hall current sensor, a piezoelectric ceramic pressure sensor, and an ultrasonic acoustic emission sensor are deployed at the P-type product melt spot welding station. The synchronous sampling frequency of all sensors is set to 10 kHz to acquire analog electrical signals containing spot welding current pulses, electrode pressure, and acoustic emission waves, and perform analog-to-digital conversion. The precise time point at which the rising edge of the spot welding current pulse crosses a preset current threshold is marked as the spot welding current pulse trigger time. Using this trigger time as the reference zero point, a first time-domain data segment with a duration of 50 ms is extracted in reverse along the time axis as the pre-time-domain window, and a second time-domain data segment with a duration of 150 ms is extracted in forward along the time axis as the post-time-domain window.

[0010] The peak pressure, pressure ramp-up slope, and average pressure during the holding phase are extracted from the truncated electrode pressure time series. The acoustic emission energy integral, maximum amplitude, and signal decay time constant are extracted from the truncated melt nucleus formation time-domain signal. These extracted values ​​are then concatenated in a fixed order to form a one-dimensional feature vector. Fast Fourier Transform is used to calculate the initial phase angle of the first dominant frequency of the electrode pressure time series and the initial phase angle of the second dominant frequency of the melt nucleus formation time-domain signal. The absolute value of the difference between the two is calculated, and this absolute value is divided by the sum of the absolute value of the initial phase angle of the first dominant frequency and a preset positive constant to obtain the phase difference ratio. A preset second threshold value is set, and this second threshold value is divided by the sum of the phase difference ratio and a constant 1 to obtain the adjusted second threshold. When the phase difference ratio increases, the adjusted second threshold decreases accordingly to improve the sensitivity to the cumulative deviation triggering strong persistence of synchronization under phase deviation conditions.

[0011] In an optional embodiment, the step of extracting feature vectors by performing dual-time-domain window truncation on the time series and the time-domain signal based on the trigger time includes: Taking the triggering time as the starting reference zero point, a first time domain data segment with a duration of 50 milliseconds is extracted in reverse along the time axis as the pre-time domain window, and a second time domain data segment with a duration of 150 milliseconds is extracted in forward along the time axis as the post-time domain window. The data in the preceding time domain window and the data in the following time domain window are vertically spliced ​​or cascaded in the row direction along the time axis to obtain a dual time domain feature segment of uniform specification. The corresponding feature values ​​are extracted from the electrode pressure time series and the melt nucleus formation time domain signal in the dual time domain feature segment, and the extracted feature values ​​are spliced ​​in a fixed order to generate the feature vector.

[0012] A continuously stored, 2000×2 dual-time-domain feature extraction segment matrix is ​​generated. The initial sampling parameters and feature extraction logic are established as the basic context: based on the system's 10kHz sampling frequency and the zero-point position triggered by the absolute timestamp, the controller extracts a pre-time-domain window containing 500 discrete sampling points and a post-time-domain window containing 1500 discrete sampling points. The processing module skips the intermediate redundant variable transition mapping step and calls the built-in matrix function to perform a vertical concatenation operation along the time axis. The 500×2 pre-matrix and the 1500×2 post-matrix are concatenated along the row direction to form a 2000×2 uniform dual-time-domain feature extraction segment matrix. This dual-time-domain feature extraction segment is used to define the data organization range of the pre-time-domain window and the post-time-domain window. Subsequent feature extraction is performed according to the electrode pressure time series and the melt nucleus formation time-domain signal, respectively, and the extraction results are concatenated into a one-dimensional feature vector in a fixed order.

[0013] In an optional embodiment, the step of calculating the phase difference ratio based on the time series and the time-domain signal, and obtaining the adjusted second threshold based on the phase difference ratio, includes: The initial phase angle of the first dominant frequency of the time series and the initial phase angle of the second dominant frequency of the time domain signal are calculated using Fast Fourier Transform. Calculate the absolute value of the difference between the initial phase angle of the first main frequency and the initial phase angle of the second main frequency. Use the absolute value of the difference as the numerator and the sum of the absolute value of the initial phase angle of the first main frequency and a preset positive constant as the denominator to obtain the phase difference ratio. The phase difference ratio is summed with 1 to obtain a sum value. The preset reference second threshold value is divided by this sum value to obtain the adjusted second threshold.

[0014] In this embodiment, the adjusted second threshold is ultimately set to 3.825. The initial phase angle of the first main frequency is 1.20 radians, and the initial phase angle of the second main frequency is 0.80 radians, with an absolute difference of 0.40 radians. With a preset positive constant of 0.1, the phase difference ratio R = 0.40 / (1.20 + 0.1) = 0.307. The baseline second threshold is 5.0, so the adjusted second threshold is 5.0 / (1 + 0.307) = 3.825, and this value is written to the relevant threshold storage address.

[0015] S2, the feature vector is pushed into the asynchronous cache queue, the weight is calculated based on the absolute value of the deviation between the feature vector and the standard melt parameters and the ratio of the phase difference, the priority read index is constructed for the queue based on the weight without changing the queue timing arrangement, the memory occupancy rate of the queue is monitored, the sliding cumulative deviation matrix is ​​calculated based on the feature vector maintaining the timing arrangement, when the occupancy rate is greater than the preset first threshold or the trace of the matrix is ​​greater than the adjusted second threshold, the synchronous strong persistence mode is switched, and the reading order of the feature vector to be persisted is determined according to the priority read index in the synchronous strong persistence mode.

[0016] Configure a contiguous thread-safe circular buffer in the server system memory as an asynchronous cache queue. Store the extracted feature vectors at the tail of this asynchronous cache queue according to the data acquisition time sequence. Retrieve pre-stored standard melt parameters to construct a standard feature vector. Calculate the dimensionless absolute deviation between the currently stored feature vector and the corresponding dimensions of the standard feature vector, and sum them to obtain the global absolute deviation value D. Denote the sum of D and 1 as Y, and then... Calculate the weight value assigned to the current feature vector, where R is the phase difference ratio. Establish a singly linked list structure in system memory, completely independent of the asynchronous cache queue, as a priority index table. Each node in this priority index table records the memory address pointer and corresponding weight value of a single feature vector in the asynchronous cache queue. Use the quicksort algorithm to sort the nodes in the priority index table in descending order of weight value. During this sorting process, only the links of the pointers in the index table are changed; no movement is performed on the storage continuity or position of the feature vectors within the asynchronous cache queue to maintain the original temporal order.

[0017] The memory occupancy rate is calculated by periodically reading the asynchronous cache queue at a rate of one second, and then using the percentage of currently used memory bytes out of the total allocated memory bytes. A sliding window of length 5 is used to extract the five most recent, sequentially ordered feature vectors from the asynchronous cache queue, following a first-in, first-out (FIFO) principle. Each feature vector and its standard feature vector are then dimensionless, and each feature vector is subtracted from the standard feature vector to obtain five residual column vectors. The elements of the corresponding positions in these five residual outer product matrices are summed to construct a sliding cumulative deviation matrix. The trace of this matrix is ​​calculated by summing the values ​​of all elements on the main diagonal of the sliding cumulative deviation matrix. A first threshold of 80% is preset. The system's main controller compares the memory occupancy rate with the calculated trace of the sliding cumulative deviation matrix using logical judgment statements. If the memory occupancy rate exceeds 80% or the calculated trace of the sliding cumulative deviation matrix is ​​greater than the adjusted second threshold, a highest-level hardware interrupt signal is immediately sent to the processor bus. Suspend the background asynchronous thread currently responsible for writing, enable a global mutex lock to redirect the system data stream write path to the storage sector of the solid-state drive and enter synchronous strong persistence mode. Then, read the feature vectors that have not been written to disk in the index table in descending order of priority, and retain the timing identifier of each feature vector in the original asynchronous cache queue before performing persistent writing.

[0018] In an optional embodiment, calculating the weight based on the ratio of the absolute value of the deviation between the feature vector and the standard melt parameters and the phase difference includes: The global absolute deviation value D, and the corresponding numerical data synchronization priority weight W, are generated by calculating the absolute deviation values ​​corresponding to each feature value using the following formula: ; ; Where n is the dimension of the feature values, Let j be the value of the j-th feature in the feature vector. Let be the j-th corresponding non-zero reference value in the standard melt parameters, ln be the natural logarithm function, Y be the sum of D and 1, and R be the phase difference ratio.

[0019] In this embodiment, the calculated synchronization priority weight is 0.1951. The measured multidimensional feature data is [2100, 1550, 0.85], and the corresponding non-zero baseline value is [2000, 1500, 0.80]. The absolute deviation values ​​corresponding to the first, second, and third dimensions are 0.05, 0.033, and 0.0625, respectively, and the sum of these values ​​yields a global absolute deviation parameter D≈0.1455. Y is denoted as the sum of D and 1, i.e., Y=1.1455. Substituting R=0.307, we get W=0.307×ln(1.1455)+0.5×0.307≈0.1951. This value is used as the synchronization priority weight of the current feature vector. Here, 0.5×R is used to retain the basic synchronization priority when the parameter deviation is small but the phase difference ratio increases, so that phase abnormal data can still be preferentially included in the persistent judgment.

[0020] In an optional embodiment, constructing a priority read index for the queue based on the weight without changing the queue's temporal order includes: Extract the values ​​of the weights corresponding to all currently existing feature vectors in the queue; The extracted values ​​are compared in descending order from largest to smallest. An independent priority read index is constructed according to the comparison sorting sequence. The timing logic pointers of each feature vector storage node in the queue remain unchanged, so that the feature vector with the largest weight value is placed first in the priority read index. The priority read index is used as the basis for determining the read order of the feature vectors to be persisted in the synchronous strong persistence mode.

[0021] In this embodiment, the weight values ​​corresponding to nodes one through four in the queue are 0.12, 0.35, 0.08, and 0.21, respectively. The system sorts the nodes according to their weight values ​​from largest to smallest, resulting in 0.35 > 0.21 > 0.12 > 0.08. Therefore, the priority read index order is [node two, node four, node one, node three]. After entering the synchronous strong persistence mode, the system reads the feature vectors corresponding to nodes two, four, one, and three in sequence according to this priority read index, and writes the original acquisition time sequence number of each node into the database to ensure that high-weight data is written to disk first and the original time sequence is traceable.

[0022] In an optional embodiment, calculating the sliding cumulative deviation matrix based on the eigenvectors that maintain the temporal arrangement includes: Extract five consecutive feature vectors located at the tail of the queue at the current moment on the time axis to form a basic window data set; The sliding cumulative deviation matrix S of the current time slot is calculated based on the residual external cumulative product formula: ; in, For the k-th feature column vector in the basic window data set, For the corresponding expected state feature column vector, for and The residual column vector obtained by subtraction, Let be the transpose row vector of the residual column vector.

[0023] In this embodiment, the sliding cumulative deviation matrix generated by the residual external cumulative product operation is: Define a dimensionless column vector of desired state features. Then, extract five consecutive feature vectors to calculate the residual column vector. Using the first residual column vector... For example, multiplying it by its own transpose yields the residual outer product matrix. The sliding cumulative deviation matrix is ​​obtained by summing the five residual outer product matrices one by one. Therefore, the trace of the sliding cumulative deviation matrix is ​​0.06 + 0.10 = 0.16, and this trace value is used to compare with the adjusted second threshold.

[0024] S3, when it is determined that there is a data frame loss, a compensation matrix is ​​constructed based on the electrode pressure phase offset of the adjacent valid frames. The timestamp of the lost data frame is interpolated and re-indexed according to the compensation matrix to generate a dataset. The window truncation length of the subsequent production data is scaled based on the main eigenvalue of the compensation matrix and then archived to the database.

[0025] In the data parsing phase, the system timestamp and incrementing data frame sequence number contained in the production data header are read frame by frame. When the logic verification module detects that the sequence numbers of two adjacent valid data frames are not continuous, and the time interval between the system timestamps is greater than or equal to twice the preset data sampling period, it is determined that there is a data frame loss during system data transmission. The two consecutive valid frames immediately before and immediately after the time scale of the lost data frame are extracted as four adjacent boundary frames. Frequency domain phase extraction is performed on the electrode pressure sequence of each four adjacent boundary frame to obtain the corresponding measured phase of the electrode pressure sequence.

[0026] The absolute value of the difference between the measured phase of the electrode pressure sequence and the theoretical phase of the ideal period in each of the four adjacent boundary frames is divided by the theoretical phase of the ideal period to obtain four electrode pressure phase offsets. These four electrode pressure phase offsets are then sequentially filled into the main diagonal positions of an initial 4×4 matrix to construct a 4×4 diagonal compensation matrix. The theoretical phase of the ideal period refers to the reference expected value of the initial phase of the dominant frequency of the electrode pressure time sequence under standard fault-free spot welding conditions. Specifically, when the system retrieves standard melt parameters for normal spot welding, a standard electrode pressure time sequence is acquired. The first dominant frequency initial phase angle of this standard electrode pressure time sequence is calculated using the same fast Fourier transform as in step S1 above, and the obtained value is taken as the theoretical phase of the ideal period.

[0027] When interpolating and reindexing the timestamps of lost data frames, the timestamps corresponding to the four adjacent boundary frames are converted into time offsets relative to the trigger time of the spot welding current pulse, forming a time offset column vector. The dimensionless electrode pressure phase offset in the compensation matrix is ​​used as a time offset correction coefficient to weight and correct the boundary time offsets in the time offset column vector, resulting in a phase-corrected boundary time offset. Then, based on the sequence position of the lost data frame between the preceding and following valid frames, linear interpolation is performed on the phase-corrected boundary time offset to obtain the estimated timestamp of the lost data frame. Using this estimated timestamp as the baseline input variable, a cubic spline interpolation algorithm is used to generate a feature vector simulation compensation value corresponding to the estimated timestamp, combined with the feature vectors of the preceding and following valid frames. The estimated timestamp and the feature vector simulation compensation value are then combined to fill in the gaps in the original asynchronous buffer queue sequence, generating a time-continuous complete dataset.

[0028] In this application, the principal eigenvalue of the compensation matrix refers to the largest singular value extracted after calculating all singular values ​​of the compensation matrix using singular value decomposition (SVD). By performing SVD on the 4×4 diagonal compensation matrix, a corresponding sequence of singular values ​​is obtained, and the largest singular value is selected as the principal eigenvalue. The initial system window length, expressed in terms of the number of sampling points, is used. Multiplied by e The result is raised to the power of the power and then rounded down to obtain the window's integer value. ,in, The primary feature value is used. The window is shaped to obtain the value. Update the baseline window width register parameter for the production data collected in the next frame. Then, insert the previously generated complete dataset into the production history data storage table of the relational database using a batch transaction commit to complete the archiving operation.

[0029] In an optional embodiment, constructing the compensation matrix based on the electrode pressure phase offset of adjacent valid frames includes: Extract the two consecutive valid frames immediately before and immediately after the time scale of the lost data frame as the four adjacent boundary frames; The absolute value of the difference between the measured phase and the theoretical phase of the ideal period of the electrode pressure sequence of each of the four adjacent boundary frames is divided by the theoretical phase of the ideal period, and the ratio obtained is used as a scalar value. , , , The compensation matrix is ​​formed by filling in the positions along the main diagonal from the top left to the bottom right of the initial square matrix, where, , , , The values ​​are the phase offset values ​​of the four electrode pressures determined based on the occurrence timing.

[0030] In this embodiment, the generated compensation matrix is, for example, Taking the previous boundary frame as an example, its measured phase of the electrode pressure sequence is 3.29 radians, and the theoretical phase of the ideal period is 3.14 radians. Therefore, the corresponding dimensionless electrode pressure phase offset is... Similarly, we can obtain... , , The aforementioned values ​​are then sequentially filled into the main diagonal positions of the 4×4 initial matrix to obtain the compensation matrix. The phase offset compensation matrix corresponding to the boundary frame is as follows: Figure 2 As shown.

[0031] In an optional embodiment, scaling the window truncation length of subsequent production data based on the principal eigenvalues ​​of the compensation matrix includes: All singular values ​​of the compensation matrix are calculated using the singular value decomposition method, and the largest singular value with the highest numerical value is extracted as the principal eigenvalue. Get window-shaped values ​​in units of sampling points. ; in, This refers to the original window length value, expressed in terms of the number of sampling points, corresponding to the baseline communication state of the spot welding process. Here, e is the principal eigenvalue, and e is the natural constant. The reference window width register parameter, which overwrites the production data collected in the next frame, is updated using the window reshaping value.

[0032] In this embodiment, the principal eigenvalue of the compensation matrix is ​​0.0512. The initial window length is set as follows. =2000, thus obtaining =2105. Use 2105 as the window reshaping value for the production data collected in the next frame, and write it into the baseline window width register parameter.

[0033] In the second embodiment, the present invention also proposes an intelligent production management system for melt spot welding of P-type products, comprising the following modules: The acquisition module is used to acquire production data of the spot welding process, including the spot welding current pulse trigger time, electrode pressure time series and weld nugget formation time domain signal. Based on the trigger time, the time series and the time domain signal are subjected to dual time domain window truncation to extract feature vectors. The phase difference ratio is calculated according to the time series and the time domain signal, and the adjusted second threshold is obtained based on the phase difference ratio. The module is used to push the feature vector into an asynchronous cache queue, calculate the weight based on the absolute value of the deviation between the feature vector and the standard melt parameters and the ratio of the phase difference, build a priority read index for the queue based on the weight without changing the queue's temporal order, monitor the memory usage of the queue, calculate the sliding cumulative deviation matrix based on the feature vector while maintaining the temporal order, and switch to synchronous strong persistence mode when the memory usage exceeds a preset first threshold or the trace of the matrix exceeds the adjusted second threshold. In the synchronous strong persistence mode, the reading order of the feature vectors to be persisted is determined according to the priority read index. The generation module is used to construct a compensation matrix based on the electrode pressure phase offset of adjacent valid frames when it is determined that a data frame is lost. It then interpolates and re-indexes the timestamps of the lost data frames according to the compensation matrix to generate a dataset. Finally, it scales the window length of subsequent production data based on the principal eigenvalues ​​of the compensation matrix and archives it into the database.

[0034] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. An intelligent production management method for melt spot welding of P-type products, characterized in that, Includes the following steps: Production data of the spot welding process is collected, including the spot welding current pulse trigger time, electrode pressure time series and weld nugget formation time domain signal. Based on the trigger time, the time series and the time domain signal are subjected to dual time domain window truncation to extract feature vectors. The phase difference ratio is calculated according to the time series and the time domain signal, and the adjusted second threshold is obtained based on the phase difference ratio. The feature vector is pushed into an asynchronous cache queue. A weight is calculated based on the absolute value of the deviation between the feature vector and the standard melt parameters and the ratio of the phase difference. A priority read index is constructed for the queue based on the weight without changing the queue's temporal order. The memory occupancy rate of the queue is monitored. A sliding cumulative deviation matrix is ​​calculated based on the feature vectors that maintain the temporal order. When the occupancy rate is greater than a preset first threshold or the trace of the matrix is ​​greater than the adjusted second threshold, the system switches to synchronous strong persistence mode. In the synchronous strong persistence mode, the read order of the feature vectors to be persisted is determined according to the priority read index. When it is determined that a data frame is lost, a compensation matrix is ​​constructed based on the electrode pressure phase offset of the adjacent valid frames. The timestamps of the lost data frames are interpolated and re-indexed according to the compensation matrix to generate a dataset. The window truncation length of subsequent production data is scaled based on the principal eigenvalue of the compensation matrix and then archived to the database.

2. The method according to claim 1, characterized in that, The step of extracting feature vectors by performing dual-time-domain window truncation on the time series and the time-domain signal based on the trigger time includes: Taking the triggering time as the starting reference zero point, a first time domain data segment with a duration of 50 milliseconds is extracted in reverse along the time axis as the pre-time domain window, and a second time domain data segment with a duration of 150 milliseconds is extracted in forward along the time axis as the post-time domain window. The data in the preceding time domain window and the data in the following time domain window are vertically spliced ​​or cascaded in the row direction along the time axis to obtain a dual time domain feature segment of uniform specification. The corresponding feature values ​​are extracted from the electrode pressure time series and the melt nucleus formation time domain signal in the dual time domain feature segment, and the extracted feature values ​​are spliced ​​in a fixed order to generate the feature vector.

3. The method according to claim 1, characterized in that, The step of calculating the phase difference ratio based on the time series and the time-domain signal, and obtaining the adjusted second threshold based on the phase difference ratio, includes: The initial phase angle of the first dominant frequency of the time series and the initial phase angle of the second dominant frequency of the time domain signal are calculated using Fast Fourier Transform. Calculate the absolute value of the difference between the initial phase angle of the first main frequency and the initial phase angle of the second main frequency. Use the absolute value of the difference as the numerator and the sum of the absolute value of the initial phase angle of the first main frequency and a preset positive constant as the denominator to obtain the phase difference ratio. The phase difference ratio is summed with 1 to obtain a sum value. The preset reference second threshold value is divided by this sum value to obtain the adjusted second threshold.

4. The method according to claim 1, characterized in that, The step of calculating the weight based on the absolute value of the deviation between the feature vector and the standard melt parameters and the ratio of the phase difference includes: The global absolute deviation value D, and the corresponding numerical data synchronization priority weight W, are generated by calculating the absolute deviation values ​​corresponding to each feature value using the following formula: ； W = R × lnY + 0.5 × R; Where n is the dimension of the feature values, Let j be the value of the j-th feature in the feature vector. Let be the j-th corresponding non-zero reference value in the standard melt parameters, ln be the natural logarithm function, Y be the sum of D and 1, and R be the phase difference ratio.

5. The method according to claim 4, characterized in that, The step of constructing a priority read index for the queue based on the weight without changing the queue's temporal order includes: Extract the values ​​of the weights corresponding to all currently existing feature vectors in the queue; The extracted values ​​are compared in descending order from largest to smallest. An independent priority read index is constructed according to the comparison sorting sequence. The timing logic pointers of each feature vector storage node in the queue remain unchanged, so that the feature vector with the largest weight value is placed first in the priority read index. The priority read index is used as the basis for determining the read order of the feature vectors to be persisted in the synchronous strong persistence mode.

6. The method according to claim 1, characterized in that, The calculation of the sliding cumulative deviation matrix based on the eigenvectors that maintain the temporal arrangement includes: Extract five consecutive feature vectors located at the tail of the queue at the current moment on the time axis to form a basic window data set; The sliding cumulative deviation matrix S of the current time slot is calculated based on the residual external cumulative product formula: ; in, For the k-th feature column vector in the basic window data set, For the corresponding expected state feature column vector, for and The residual column vector obtained by subtraction, Let be the transpose row vector of the residual column vector.

7. The method according to claim 1, characterized in that, The compensation matrix is ​​constructed based on the electrode pressure phase offset of adjacent valid frames, including: Extract the two consecutive valid frames immediately before and immediately after the time scale of the lost data frame as the four adjacent boundary frames; The absolute value of the difference between the measured phase and the theoretical phase of the ideal period of the electrode pressure sequence of each of the four adjacent boundary frames is divided by the theoretical phase of the ideal period, and the ratio obtained is used as a scalar value. , , , The compensation matrix is ​​formed by filling in the positions along the main diagonal from the top left to the bottom right of the initial square matrix, where, , , , The values ​​are the phase offset values ​​of the four electrode pressures determined based on the occurrence timing.

8. The method according to claim 1, characterized in that, The scaling of the window truncation length for subsequent production data based on the principal eigenvalues ​​of the compensation matrix includes: All singular values ​​of the compensation matrix are calculated using the singular value decomposition method, and the largest singular value with the highest numerical value is extracted as the principal eigenvalue. Get window-shaped values ​​in units of sampling points. ; in, This refers to the original window length value, expressed in terms of the number of sampling points, corresponding to the baseline communication state of the spot welding process. Here, e is the principal eigenvalue, and e is the natural constant. The reference window width register parameter, which overwrites the production data collected in the next frame, is updated using the window reshaping value.

9. An intelligent production management system for melt spot welding of P-type products, characterized in that, Includes the following modules: The acquisition module is used to acquire production data of the spot welding process, including the spot welding current pulse trigger time, electrode pressure time series and weld nugget formation time domain signal. Based on the trigger time, the time series and the time domain signal are subjected to dual time domain window truncation to extract feature vectors. The phase difference ratio is calculated according to the time series and the time domain signal, and the adjusted second threshold is obtained based on the phase difference ratio. The module is used to push the feature vector into an asynchronous cache queue, calculate the weight based on the absolute value of the deviation between the feature vector and the standard melt parameters and the ratio of the phase difference, build a priority read index for the queue based on the weight without changing the queue's temporal order, monitor the memory usage of the queue, calculate the sliding cumulative deviation matrix based on the feature vector while maintaining the temporal order, and switch to synchronous strong persistence mode when the memory usage exceeds a preset first threshold or the trace of the matrix exceeds the adjusted second threshold. In the synchronous strong persistence mode, the reading order of the feature vectors to be persisted is determined according to the priority read index. The generation module is used to construct a compensation matrix based on the electrode pressure phase offset of adjacent valid frames when it is determined that a data frame is lost. It then interpolates and re-indexes the timestamps of the lost data frames according to the compensation matrix to generate a dataset. Finally, it scales the window length of subsequent production data based on the principal eigenvalues ​​of the compensation matrix and archives it into the database.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-8.

Images