Low pressure casting machine remote fault monitoring system

Abstract

The application discloses a low-pressure casting machine remote fault monitoring system and relates to the technical field of fault monitoring. A collection and processing module collects smelting temperature, mold temperature, pressure maintaining pressure, hydraulic pressure, liquid lifting speed and motor current at different sampling frequencies, arranges the parameters into a parameter vector set, performs wavelet transform noise reduction, integrates the parameters into an aligned parameter matrix based on parameter interpolation chain interpolation and marks the equipment number; a monitoring and analysis module splits the aligned parameter matrix into three sub-matrices, extracts the sub-feature vector of each sub-matrix through one-dimensional convolution of cosine and attention mechanism, determines whether the sub-matrix is faulty by using an SNN preliminary screening optimized through Q learning, inputs the sub-feature vector of the faulty sub-matrix into a corresponding branch mapping layer to determine the fault type, determines the fault level according to the normalized Euclidean distance between the sub-feature vector of the faulty sub-matrix and the corresponding group template vector and generates a fault report; and an alarm module triggers a hierarchical alarm according to the fault report, thereby improving monitoring accuracy and efficiency.

Description

Technical Field

[0001] This invention relates to the field of fault monitoring technology, specifically to a remote fault monitoring system for low-pressure casting machines. Background Technology

[0002] During the operation of a low-pressure casting machine, the stability of key parameters such as melting temperature, mold temperature, holding pressure, hydraulic pressure, liquid lifting speed, and motor current directly affects the casting quality, equipment operation safety, and production continuity. If abnormal parameters are not monitored and handled in a timely manner, it can easily lead to casting scrap, equipment failure and shutdown, or even safety accidents. Therefore, real-time and accurate monitoring of key parameters of a low-pressure casting machine is an important foundation for ensuring stable and efficient casting production and reliable equipment operation.

[0003] Existing fault monitoring technologies for low-pressure casting machines have shortcomings:

[0004] The sampling frequency was not set differently according to the response characteristics of different parameters, resulting in the omission of key change information of some parameters or redundant sampling, which increased the processing burden.

[0005] When parameters are missing, there is a lack of interpolation design that combines process constraints with parameter correlation. It is impossible to fill in the missing parameters according to process logic to achieve parameter timing alignment, resulting in the missing parameter matrix failing to accurately reflect the equipment operating status.

[0006] In the fault analysis phase, it is difficult to balance the efficiency and interpretability of feature extraction, and there is a lack of a fast and effective fault screening mechanism, resulting in low accuracy and efficiency in fault type identification and level determination. Summary of the Invention

[0007] This invention addresses the shortcomings of existing technologies by proposing a remote fault monitoring system for low-pressure casting machines. It adapts parameter characteristics through differentiated sampling and combines them with precise noise reduction using the db4 wavelet basis. Based on the process flow, it constructs a parameter interpolation chain to achieve parameter temporal synchronization. Through the collaborative efforts of cosine one-dimensional convolution, attention mechanism, SNN initial screening, and branch mapping structure, it achieves efficient and accurate fault analysis, thereby improving the accuracy and efficiency of fault monitoring for low-pressure casting machines.

[0008] The technical solution to achieve the purpose of this invention is as follows:

[0009] The remote fault monitoring system for low-pressure casting machines includes a data acquisition and processing module, a monitoring and analysis module, and an alarm module.

[0010] The data acquisition and processing module collects smelting temperatures at different sampling frequencies. Mold temperature Holding pressure Hydraulic pressure Liquid lifting rate and motor current In each cycle, the parameters are organized into a set of parameter vectors and noise is filtered out using wavelet transform. Based on the parameter interpolation chain, a model is selected for each parameter in sequence for interpolation. The resulting alignment parameter matrix is ​​integrated for each cycle and the equipment number is marked. Among them, the parameter interpolation chain can be pre-constructed by performing grey relational analysis on the parameters located at the co-occurrence time in the historical cycle and performing multi-dimensional screening based on process flow, correlation degree and stability.

[0011] The monitoring and analysis module splits the alignment parameter matrix of each period into three sub-matrices based on the grouping. It extracts the temporal and spatial features of each sub-matrix through cosine one-dimensional convolution and attention mechanism, and integrates them to generate the sub-feature vector of each sub-matrix. It uses an SNN optimized by Q-learning to screen and judge whether each sub-matrix is ​​faulty. If there are no faults, it remains silent until the next period. In other cases, the sub-feature vector of the faulty sub-matrix is ​​mapped to the branch fault probability vector through the corresponding branch mapping layer to determine the fault type. The fault level is determined based on the normalized Euclidean distance between the sub-feature vector of the faulty sub-matrix and the group template vector of the corresponding group. The fault report is generated by integrating the equipment number, fault type and fault level.

[0012] The alarm module implements a tiered alarm mechanism based on fault reports, triggering different alarm methods according to the fault level in each cycle.

[0013] Specifically, the acquisition and processing module is in the first... The period is based on the sampling frequency to collect the smelting temperature. and mold temperature The holding pressure was collected at twice the sampling frequency. Hydraulic pressure and liquid lifting speed Motor current is collected at 4 times the sampling frequency. , obtained the Periodic melting temperature vector Mold temperature vector Pressure holding vector Hydraulic vector acceleration vector and current vector , combine to construct the first Periodic parameter vector set .

[0014] Furthermore, the acquisition and processing module utilizes wavelet transform based on the db4 wavelet basis for noise reduction to... Periodic parameter vector set The parameter vector formed by each parameter in the vector Perform a 5-level progressive decomposition to generate the 1st level. Low-frequency coefficient vector of the 5th layer of the cycle And the 5th layer high-frequency coefficient vector, will the first Absolute median difference of the first layer of the cycle Divide by 0.6745 to get the first Periodic noise standard deviation , among which, the Absolute median difference of the first layer of the cycle equal to the High-frequency coefficient vector of the first layer of the cycle Each high-frequency coefficient in the middle and the first High-frequency coefficient vector of the first layer of the cycle the median of The median of the absolute differences and These represent taking the absolute median difference and the median of the vector, respectively. Periodic parameter vector The square root of the natural logarithm of the dimension multiplied by times the Periodic noise standard deviation , obtained the The periodic Donogh threshold And for the first Any high-frequency coefficient in the 5-layer high-frequency coefficient vector of the periodicity Perform soft thresholding when the high-frequency coefficient The absolute value is greater than the first The periodic Donogh threshold At that time, the high frequency coefficients The absolute value and the first The periodic Donogh threshold The difference multiplied by the high frequency coefficient The sign function value is used to obtain the corresponding high-frequency noise reduction coefficients. In other cases, the corresponding high-frequency noise reduction coefficient will be... Set to 0, where the sign function values ​​corresponding to negative numbers, 0, and positive numbers are -1, 0, and 1, respectively. After soft thresholding, the th... The periodic 5-layer noise reduction high-frequency coefficient vector, using the db4 wavelet basis to the first Low-frequency coefficient vector of the 5th layer of the cycle With the The high-frequency coefficient vectors of the five-layer noise reduction cycle are reconstructed by inverse transformation to generate the first... Periodic pure parameter vector .

[0015] Furthermore, when the acquisition and processing module uses the db4 wavelet basis for progressive decomposition, it will... Periodic parameter vector Considered as the first Low-frequency coefficient vector of the 0th layer of the cycle Using low-pass and high-pass filters respectively with the first The low-frequency coefficient vectors of each layer in the periodic cycle are convolved and simultaneously downsampled by a factor of 2 to generate the first... The low-frequency coefficient vector and high-frequency coefficient vector of the next layer in the cycle are repeated cyclically until the first layer is obtained. Low-frequency coefficient vector of the 5th layer of the cycle With high frequency coefficient vector After completing the progressive decomposition, when using the db4 wavelet basis for inverse transform reconstruction, the first... The low-frequency and high-frequency coefficient vectors of each layer in the period are upsampled by a factor of 2, transposed and convolved with the low-pass and high-pass filters respectively, and then summed to obtain the th... The low-frequency coefficient vector of the previous layer is reconstructed until the 0th layer is reached, generating the 1st layer. Periodic pure parameter vector .

[0016] Furthermore, the acquisition and processing module extracts six parameters from the parameter vector set of historical periods for all co-occurrence moments, and obtains a standard collinear parameter matrix by splicing and integrating them and standardizing them with Z-Score. The co-occurrence time refers to the acquisition time of the six parameters simultaneously. The standard collinearity parameter matrix is ​​calculated based on grey relational analysis. The Middle Line and number The correlation between the two parameters corresponding to the row , No. Line and number The correlation between the two parameters corresponding to the row equal to the Line and number The two parameters corresponding to the row are in The sum of the temporal correlations at each co-occurrence moment is determined by the standard collinearity parameter matrix. The Middle Standard co-occurrence parameter vector of rows Calculate the global minimum difference and global maximum difference among all standard collinear parameter differences in the standard co-occurrence parameter difference vector of each of the remaining standard collinear parameter vectors, and calculate the first... Standard co-occurrence parameter vector of rows With the Standard co-occurrence parameter vector of rows In the Co-occurrence moments The standard parameter difference is calculated by dividing the sum of the global minimum difference and 0.5 times the global maximum difference by the first parameter. Co-occurrence moments The standard collinearity parameter matrix is ​​obtained by summing the standard parameter difference with 0.5 times the global maximum difference. The Middle Line and number The two parameters corresponding to the row are in the first row. Co-occurrence moments Time correlation , , , Using six parameters as nodes, any two nodes are connected by undirected edges, and the degree of association between the two parameters corresponding to the two nodes is assigned to each node, thus constructing a parametric undirected graph.

[0017] Furthermore, the data acquisition and processing module sets the melting temperature based on the process flow. Pointing to mold temperature First directional constraint, setting motor current Pointing to the liquid rise rate The second directional constraint sets the liquid lifting rate. Indicating holding pressure and hydraulic pressure Third-party constraint, set hydraulic pressure Indicating holding pressure The fourth direction constraint is set to maintain the pressure. Pointing to mold temperature The fifth directional constraint is applied. Based on this constraint, undirected edges between corresponding nodes in the parametric undirected graph are set as direct directed edges. Based on the principle of flow connectivity, when any node reaches another node via a finite number of non-repeating direct directed edges, the undirected edge between any node and another node is set as an indirect directed edge pointing from one node to the other. Remaining undirected edges are pruned. Multi-dimensional filtering is performed based on correlation and stability thresholds, deleting directed edges with correlation less than or equal to the correlation threshold. For any remaining directed edge, the standard co-occurrence parameter vector set of the two nodes of the directed edge is extracted and decomposed into... A set of standard co-occurrence parameter segments, calculated The ratio of the standard deviation to the mean of the segment correlation degree of each standard co-occurrence parameter segment group is used to obtain the coefficient of variation of the directed edges. Directed edges with coefficients of variation greater than the stability threshold are retained to construct a parameter interpolation chain. The parameter interpolation chain includes a first interpolation sub-chain and a second interpolation sub-chain. The first interpolation sub-chain is determined by the melting temperature. Pointing to mold temperature The second interpolation sub-chain is powered by the motor current. Pointing to the liquid rise rate And simultaneously point to the holding pressure. and hydraulic pressure .

[0018] Furthermore, the acquisition and processing module references the interpolation chain and uses different models to complete the interpolation except for the motor current. The other 5 parameters in the first The value of each missing acquisition moment in the cycle is used to generate the first... Periodic alignment of melting temperature vector Alignment of mold temperature vector Alignment of acceleration vector Alignment and pressure holding vector and aligned hydraulic vector , and the Periodic current vector spliced ​​together as the first periodic alignment parameter matrix This includes the following steps:

[0019] For the first interpolation subchain, based on the... Periodic melting temperature vector The existing melting temperature is fitted with a cubic spline to obtain the first... The periodic melting temperature change curve is calculated, and the supplementary melting temperature for each missing data acquisition moment is calculated to generate the first... Periodic alignment of melting temperature vector Using a time series prediction network to extract the first The melting temperature or supplementary melting temperature at each acquisition time of the cycle, multiplied by the melting temperature. With mold temperature The degree of correlation between them is obtained. The weighted melting temperature at each acquisition moment in the cycle is used to mark the mask values ​​of each missing acquisition moment and the non-missing acquisition moment as 0 and 1 respectively. Periodic mold temperature vector For each missing acquisition moment, 0 is interpolated as the corresponding mold temperature, and the data is spliced ​​together. The weighted melting temperature, mask value, and mold temperature at each acquisition moment in the cycle are used to obtain the corresponding mold feature vector, which is then input into the Long Short-Term Memory layer. Based on the mask value of each acquisition moment (1 or 0), the system selects to jump to the next acquisition moment or to stitch together the previous one. The mold feature vector at each acquisition time of the period and the first The mold temperature at the previous acquisition time or the supplementary mold temperature is mapped to obtain the first cycle. The supplementary mold temperature at each acquisition moment in the cycle is used to complete the generation of the first... Periodic alignment of the mold temperature vector ;

[0020] For the second patch chain, the liquid rise rate Equal to output flow Divide by the area of ​​the hydraulic cylinder Output flow Equal to 1000 times the output power Divide by the pump inlet and outlet pressure difference With pump efficiency The product of the two, the output power equal times the rated operating voltage Motor current With power factor The product of these factors is used to construct a liquid-lifting mapping model and calculate the product of the first and second factors in sequence. Periodic acceleration vector The replenishment rate of fluid at each missing data collection moment constitutes the first... Periodic alignment acceleration vector Bernoulli's equation is used to establish the associated holding pressure. With the liquid rise rate The pressure holding mapping model, pressure holding It equals the constant coefficient determined by the molten metal and the riser tube multiplied by the riser velocity. square plus atmospheric pressure Calculate the number in sequence Periodic holding pressure vector The supplementary holding pressure at each missing acquisition moment constitutes the first... Periodic alignment holding pressure vector Using a time series prediction network, the first... Periodic alignment acceleration vector The liquid rise rate or supplemental liquid rise rate at each acquisition time is multiplied by the liquid rise rate. With hydraulic pressure The degree of correlation between them is obtained. The weighted liquid rise rate at each acquisition moment in the cycle will be the first... Periodic hydraulic vector The mask values ​​for missing and non-missing acquisition times are marked as 0 and 1, respectively, in the first... Periodic hydraulic vector Interpolate 0 as the corresponding hydraulic pressure at the acquisition time in the data acquisition phase to construct the first... The hydraulic feature vector at each acquisition time of the cycle is input into the long short-term memory layer, based on the first... The mask value decision at each acquisition time point in the cycle directly jumps to the next acquisition time point or maps the output. The supplementary hydraulic pressure at each acquisition moment in the cycle is used to obtain the first... Periodic alignment of hydraulic vectors .

[0021] The monitoring and analysis module will determine the first based on physical attributes. periodic alignment parameter matrix Alignment of melting temperature vectors in and Alignment Mold Temperature Vector Alignment and pressure holding vector and aligned hydraulic vector Alignment of acceleration vector and current vector The sub-matrices are concatenated into temperature, pressure, and dynamic sub-matrices, respectively. Cosine one-dimensional convolution (replacing the linear convolution kernel with a cosine kernel) and an attention mechanism are then used to process each of the three sub-matrices, yielding corresponding sub-temporal feature matrices and sub-space feature vectors. The sub-temporal feature matrix of each sub-matrix is ​​flattened and concatenated with its corresponding sub-space feature vector to obtain the... The sub-eigenvectors of each submatrix in the periodicity.

[0022] Furthermore, the monitoring and analysis module constructs a corresponding dedicated reward function for each submatrix using the Q-learning algorithm. The false positive rate under fault-free conditions and the false negative rate under fault conditions are used as negative penalties for each submatrix, while the accuracy of the SNN fault screening for each submatrix is ​​used as a positive reward. Through multiple rounds of iterative training on historical condition samples, the SNN decision threshold for each submatrix is ​​dynamically optimized to obtain the impulse mutation threshold for each submatrix. The sub-feature vectors of each submatrix and the actual feedback results after fault determination are input into the reinforcement learning framework for each period. The Q-value table for each submatrix is ​​updated incrementally online to obtain the impulse mutation threshold for each submatrix in the next period. Based on the SNN initial screening, the... Half the difference between the maximum and minimum sub-eigenvalues ​​of the sub-eigenvectors of each submatrix in the periodicity is used as the pulse excitation threshold for each submatrix. The magnitude of each sub-eigenvalue in the sub-eigenvectors of each submatrix is ​​compared with the pulse excitation threshold to set pulse markers. A pulse sequence for each submatrix is ​​generated, and the ratio of the standard deviation to the mean of the pulse sequence is calculated to obtain the pulse variation coefficient for each submatrix. The pulse variation coefficient of each submatrix is ​​compared with the value of the pulse sequence in the periodicity ... The pulse variation threshold of the period is used to determine the first Is each submatrix of the periodicity faulty?

[0023] Furthermore, the monitoring and analysis module is based on the first... Grouping of the periodic fault submatrix, the first The sub-feature vectors of the periodic fault sub-matrix are input into the corresponding branch mapping layer, and mapped to the corresponding branch fault probability vectors through two fully connected layers. The output layer then identifies the branch fault types corresponding to branch fault probabilities greater than a probability threshold. This is based on the... Extract the corresponding group template vector from the grouped fault submatrix of the periodicity, and calculate the group template vector and the... The normalized Euclidean distance of the sub-eigenvectors of the periodic fault submatrix is ​​used to determine the corresponding fault level based on the fault level interval to which the normalized Euclidean distance belongs.

[0024] Compared with existing technologies, this invention adopts differentiated sampling frequencies to adapt to different parameter characteristics, combines a 5-layer progressive decomposition based on the db4 wavelet basis with soft thresholding based on the Donohue threshold to achieve parameter denoising, and adaptively balances noise suppression and parameter detail preservation. It constructs an undirected parameter graph by analyzing historical co-occurring parameters through grey relational analysis, transforms it into a directed graph by combining it with the low-pressure casting process flow, and forms a parameter interpolation chain with a determined interpolation order after screening by relational and stability thresholds. It specifically employs cubic spline fitting, a liquid-lifting mapping model, and a time-series prediction network to fill in missing parameters, achieving parameter time-series alignment that fits the process logic. The aligned parameter matrix is ​​split into temperature, pressure, and dynamic sub-matrices. Temporal features are extracted through cosine one-dimensional convolution that balances computational efficiency and feature interpretability, and spatial features are extracted using an attention mechanism to generate sub-feature vectors. A Q-learning-optimized SNN is used for initial screening to quickly determine faults, and the corresponding branch mapping layer of the sub-matrix is ​​input to accurately anchor the fault type. The fault level is adaptively determined based on the normalized Euclidean distance between the sub-feature vectors and the group template vector, improving the efficiency and accuracy of fault judgment. Attached Figure Description

[0025] Figure 1 This is a structural diagram of a remote fault monitoring system for low-pressure casting machines.

[0026] Figure 2 The flowchart shows the progressive decomposition process based on the db4 wavelet basis.

[0027] Figure 3 This is a schematic diagram of a parametric undirected graph, a parametric pruned directed graph, and a parametric interpolation chain.

[0028] Figure 4 This is a flowchart of the initial screening process for SNN. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0030] like Figure 1 As shown, the present invention discloses a remote fault monitoring system for low-pressure casting machines, including a data acquisition and processing module, a monitoring and analysis module, and an alarm module;

[0031] The data acquisition and processing module continuously acquires the melting temperature at different sampling frequencies. Mold temperature Holding pressure Hydraulic pressure Liquid lifting rate and motor current In each cycle, a parameter vector set is generated and noise is filtered out using wavelet transform. The timing interpolation order is determined by referring to the parameter interpolation chain, and the corresponding model is selected for timing synchronous interpolation for each parameter. The alignment parameter matrix of each cycle is spliced ​​and the equipment number is marked. The matrix is ​​sent to the monitoring and analysis module via the MQTT protocol. Among them, the parameter interpolation chain can be pre-constructed by performing grey relational analysis on the parameters located at the co-occurrence time in the historical cycle and performing multi-dimensional screening based on the process flow, correlation degree and stability.

[0032] The monitoring and analysis module receives an alignment parameter matrix with labels in each cycle and splits it into three sub-matrices based on the grouping of six parameters. It extracts and generates the sub-temporal feature matrix of each sub-matrix through cosine one-dimensional convolution, generates the sub-space feature vector of each sub-matrix through an attention mechanism, integrates them to generate the sub-feature vector of each sub-matrix, and uses an SNN optimized by Q-learning to quickly evaluate whether each sub-matrix is ​​faulty. If all three sub-matrices are fault-free, it waits to receive the alignment parameter matrix with labels in the next cycle. If a sub-matrix is ​​faulty, the sub-feature vector of the faulty sub-matrix is ​​input into the corresponding branch mapping layer, outputs the branch fault probability vector, and determines the fault type based on the probability threshold. It determines the fault level based on the normalized Euclidean distance between the sub-feature vector of the faulty sub-matrix and the group template vector of the corresponding group under normal operating conditions. It integrates the equipment number, fault type and fault level to generate a fault report and delivers it to the alarm module.

[0033] Upon receiving a fault report for each cycle, the alarm module executes a tiered alarm mechanism, matching the fault level for each cycle to trigger an alarm. Level 1 faults are triggered by pushing an alarm pop-up via the WebSocket protocol. Level 2 faults are triggered by calling the SMS gateway, sending an alarm SMS via the API interface, and an alarm email via the SMTP protocol. Level 3 faults, in addition to the alarm methods used for Level 2 faults, utilize AI voice to generate an alarm voice message, initiate a voice call reminder via the VoIP protocol, and issue a shutdown command to the low-pressure casting machine via the MQTT protocol. The alarm pop-up, alarm SMS, alarm email, and alarm voice messages all include the device number, fault type, and fault level.

[0034] Specifically, the acquisition and processing module is in the first... The period is based on the sampling frequency to collect the smelting temperature. and mold temperature The holding pressure was collected at twice the sampling frequency. Hydraulic pressure and liquid lifting speed Motor current is collected at 4 times the sampling frequency. , obtained the Periodic melting temperature vector Mold temperature vector Pressure holding vector Hydraulic vector acceleration vector and current vector The details are as follows:

[0035] ,

[0036] ,

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] in, , , , , and The first The first acquisition time of the cycle The melting temperature, mold temperature, holding pressure, hydraulic pressure, liquid lifting speed, and motor current, and the first Cycle number Each collection time With the Each collection time The interval between them is equal to the reciprocal of four times the sampling frequency. , The maximum number of samples per cycle is always equal to 4 times the sampling frequency multiplied by the single cycle duration. There is a multiple relationship between the sampling frequencies of the six parameters. Except for the motor current, the other five parameters are sampled in the [number of samples in the cycle]. No corresponding data was collected at certain sampling times within the period, and the combination of the first and second sampling times was not recorded. Periodic melting temperature vector Mold temperature vector Pressure holding vector Hydraulic vector acceleration vector and current vector , construct the first Periodic parameter vector set , In this embodiment, the single cycle duration is 30 seconds and the sampling frequency is 1000 Hz.

[0042] Furthermore, the acquisition and processing module utilizes wavelet transform based on the db4 wavelet basis for noise reduction on the first... Periodic parameter vector set The parameter vector formed by each parameter in the vector Denoising is performed, the first Periodic parameter vector Including melting temperature vector Mold temperature vector Pressure holding vector Hydraulic vector acceleration vector and current vector Using the db4 wavelet basis to the first Periodic parameter vector Perform a 5-level progressive decomposition, and then... Periodic parameter vector The components of the non-common segment are allocated to different layers to generate the first layer. Low-frequency coefficient vector of the 5th layer of the cycle and 5-layer high-frequency coefficient vector , , , and ,in, For the first Cycle number High-frequency coefficient vector of the layer, number of floors The smaller the value, the higher the corresponding frequency band. Since noise is concentrated in the high-frequency region, the first... High-frequency coefficient vector of the first layer of the cycle The noise proportion is the highest, based on the first High-frequency coefficient vector of the first layer of the cycle Estimate the first Periodic noise standard deviation The details are as follows:

[0043] ,

[0044] in, Indicates the first The absolute median difference of the first layer of the cycle reflects the... High-frequency coefficient vector of the first layer of the cycle The degree of dispersion, by confirming the first High-frequency coefficient vector of the first layer of the cycle the median of , will the High-frequency coefficient vector of the first layer of the cycle Each high-frequency coefficient and the median Taking the difference and its absolute value, we get the first... High-frequency coefficient vector of the first layer of the cycle The absolute deviation of each high-frequency coefficient is the median of the absolute deviations. Absolute median difference of the first layer of the cycle 0.6745 is the standardized coefficient corresponding to the median of a normal distribution. and These represent taking the absolute median difference and the median of the vector, respectively. Periodic parameter vector The square root of the natural logarithm of the dimension multiplied by times the Periodic noise standard deviation This yields the first result that can adaptively balance noise suppression and vector detail preservation. The periodic Donogh threshold , for the Any high-frequency coefficient in the 5-layer high-frequency coefficient vector of the periodicity Perform soft thresholding to suppress noise in the high-frequency band, when the high-frequency coefficient The absolute value is greater than the first The periodic Donogh threshold At that time, the high frequency coefficients The absolute value and the first The periodic Donogh threshold The difference multiplied by the high frequency coefficient The sign function value is used to obtain the corresponding high-frequency noise reduction coefficients. When the high frequency coefficient The absolute value is less than or equal to the first The periodic Donogh threshold At that time, the corresponding high-frequency noise reduction coefficients will be... Set to 0, where, The sign function is defined as follows: negative numbers, 0, and positive numbers have sign function values ​​of -1, 0, and 1, respectively. After soft thresholding, the th... The 5-layer high-frequency coefficient vector of the periodicity is transformed into the first... 5-layer noise reduction high-frequency coefficient vector of the cycle , , , and ,in, For the first Cycle number The high-frequency coefficient vector of the noise reduction layer, due to the first layer Low-frequency coefficient vector of the 5th layer of the cycle For the first Periodic parameter vector The core information components are directly selected and retained, and the db4 wavelet basis is used to analyze the first... Low-frequency coefficient vector of the 5th layer of the cycle With the The high-frequency coefficient vectors of the five-layer noise reduction cycle are reconstructed by inverse transformation to generate the first... Periodic pure parameter vector .

[0045] like Figure 2 As shown, furthermore, the acquisition and processing module uses the db4 wavelet basis, which is an orthogonal wavelet with a fourth-order vanishing moment. When performing progressive decomposition using the db4 wavelet basis, the fourth... Periodic parameter vector Considered as the first Low-frequency coefficient vector of the 0th layer of the cycle Using low-pass and high-pass filters respectively with the first The low-frequency coefficient vectors of each layer in the periodic cycle are convolved and simultaneously downsampled by a factor of 2 to generate the first... The low-frequency coefficient vector and high-frequency coefficient vector of the next layer in the cycle are repeated cyclically until the first layer is obtained. Low-frequency coefficient vector of the 5th layer of the cycle With high frequency coefficient vector To complete the progressive decomposition, similarly, when using the db4 wavelet basis for inverse transform reconstruction, the th... The low-frequency and high-frequency coefficient vectors of each layer in the period are upsampled by a factor of 2, transposed and convolved with the low-pass and high-pass filters respectively, and then summed to obtain the th... The low-frequency coefficient vector of the previous layer is reconstructed until the 0th layer is reached, generating the 1st layer. Periodic pure parameter vector When the first Periodic parameter vector For the first Periodic melting temperature vector or mold temperature vector or pressure holding vector or hydraulic vector or acceleration vector or current vector At that time, the first Periodic pure parameter vector For the first Periodic pure melting temperature vector Or pure mold temperature vector Or pure pressure holding vector Or pure hydraulic vector Or pure acceleration vector or pure current vector .

[0046] Furthermore, the acquisition and processing module confirms the co-occurrence times in the parameter vector set of historical periods and extracts the corresponding 6 parameters in sequence. The 6 parameters at the same co-occurrence time are concatenated row by row into a column vector, and then concatenated column by column according to the order of co-occurrence times to obtain the co-occurrence parameter matrix. Co-occurrence parameter matrix The dimension is , This represents the total number of co-occurrence times, where each of the six parameters is acquired simultaneously. The co-occurrence parameter matrix is ​​defined as follows. Z-score normalization is performed on each row to obtain the standard collinearity parameter matrix. Combined standard co-occurrence parameter matrix The Middle line and number Okay, we get the standard co-occurrence parameter vector set. , , Based on the principle of permutation and combination, 6 parameters can be combined to obtain 15 unique standard co-occurrence parameter vector sets. For the standard co-occurrence parameter vector set... Using grey relational analysis to calculate the first Standard co-occurrence parameter vector of rows With the Standard co-occurrence parameter vector of rows The correlation of change trends along the time series dimension yields the first... line and number The correlation between the two parameters corresponding to the row The specific formula is as follows:

[0047] ,

[0048] ,

[0049] in, For the first A co-occurring moment, For the first Standard co-occurrence parameter vector of rows With the Standard co-occurrence parameter vector of rows In the Co-occurrence moments temporal correlation and They represent the first Standard co-occurrence parameter vector of rows With the Standard co-occurrence parameter vector of rows The minimum and maximum standard co-occurrence parameter differences in the standard co-occurrence parameter difference vector. and Representing the row number respectively Find the minimum of the minimum standard co-occurrence parameter difference and the maximum of the maximum standard co-occurrence parameter difference when taking different values, that is, find the first... Standard co-occurrence parameter vector of rows Among the standard co-occurrence parameter difference vectors that are collinear with the standard parameter vectors of each other row, the global minimum standard co-occurrence parameter difference and the global maximum standard co-occurrence parameter difference are found. , , and The first Standard co-occurrence parameter vector of rows With the Standard co-occurrence parameter vector of rows In the Co-occurrence moments The standard parameters are used to calculate the correlation degree of 15 non-repeating standard co-occurrence parameter vector groups in sequence. Since each row of standard collinear parameter vectors uniquely corresponds to one parameter, the six parameters are used as nodes, and the nodes are connected by undirected edges. For each undirected edge, the corresponding standard co-occurrence parameter vector group is determined based on the two ends of the undirected edge and the corresponding correlation degree is assigned to it, thus constructing a parameter undirected graph.

[0050] like Figure 3 As shown, furthermore, the data acquisition and processing module references the process flow of a low-pressure casting machine, including the melting temperature. The initial temperature of the molten metal determines the mold temperature. Set the melting temperature Pointing to mold temperature First directional constraint, motor current The motor load determines the motor speed, which is then converted into the liquid lifting speed via the transmission mechanism. Set the motor current Pointing to the liquid rise rate Second directional constraint, liquid rise velocity The hydraulic pressure reflects the flow state of the molten metal, and the flow resistance of the molten metal directly determines the hydraulic pressure. The filling volume during the liquid raising stage needs to be matched with the holding pressure. Matching and setting the liquid lifting rate Synchronous pointing holding pressure and hydraulic pressure Third-party constraint, hydraulic pressure Insufficient holding pressure will lead to insufficient holding pressure Lower, set hydraulic pressure Indicating holding pressure Fourth direction constraint, holding pressure The temperature directly determines the solidification state of the molten metal in the mold, and the exothermic solidification process indirectly affects the mold temperature. Set the holding pressure Pointing to mold temperature The fifth directional constraint is applied. Based on this constraint, undirected edges between corresponding nodes in the parametric undirected graph are set as directly directed edges. The direction of these directly directed edges must meet the directional constraint requirements. Based on the principle of flow connectivity, when any node reaches another node via a finite number of non-repeating directly directed edges, the undirected edge between any node and another node is set as an indirect directed edge. This indirect directed edge points from one node to the other. After all directly and indirectly directed edges are set, the remaining undirected edges in the parametric undirected graph are pruned, and a parametric directed graph is constructed. The directed graph contains a circular structure. The correlation between the two endpoints of a directed edge is filtered based on a correlation threshold. When the correlation is less than or equal to the threshold, the coupling between the two endpoints of the directed edge is weak, and directed edges with correlation less than or equal to the threshold are deleted, simplifying the parametric directed graph. Directed edges with correlation greater than the threshold are retained, and a parametric directed pruning graph is constructed. For two nodes with a directed edge in the parametric directed pruning graph, the corresponding standard co-occurrence parameter vector group is extracted and split at equal intervals according to the co-occurrence time order. For each group of standard co-occurrence parameter segments, calculate the segment correlation degree of each group of standard co-occurrence parameter segments, and use... The standard deviation of the correlation between segments divided by The mean of the correlation coefficients of each segment is used to obtain the coefficient of variation of the directed edges. The coefficient of variation is used to evaluate the stability of the directed edges. Directed edges with a coefficient of variation less than or equal to the stability threshold are deleted, while directed edges with a coefficient of variation greater than the stability threshold are retained. A parameter interpolation chain is constructed. The parameter interpolation chain serves as an effective basis for subsequent time-series alignment of parameters with six different sampling frequencies within the same period, ensuring that the interpolation order meets the process requirements. In this embodiment, the finally determined parameter interpolation chain includes a first interpolation sub-chain and a second interpolation sub-chain. The first interpolation sub-chain is based on the melting temperature. Starting from the mold temperature The second interpolation subchain uses the motor current. Starting from the liquid rising speed Then, the branch structure is used to synchronously point to the holding pressure. and hydraulic pressure The correlation threshold and stability threshold are set to 0.65 and 0.15, respectively.

[0051] Furthermore, the acquisition and processing module uses the first... Periodic current vector Maximum number of samples per cycle Based on this, and referring to the interpolation order and the correlation between parameters provided by the interpolation chain, the remaining five parameters are supplemented using different models in the [missing information - likely a specific context]. The value at each missing acquisition moment in the cycle is used to achieve temporal interpolation that matches the temporal trend of the parameters and the physical correlation between the parameters, generating the first... Periodic alignment of melting temperature vector Alignment of mold temperature vector Alignment of acceleration vector Alignment and pressure holding vector and aligned hydraulic vector , and the Periodic current vector By concatenating the rows, we get the first... periodic alignment parameter matrix This includes the following steps:

[0052] For the first interpolation subchain, based on the... Periodic melting temperature vector The melting temperature at each acquisition time was used to obtain the th melting temperature using cubic spline fitting. The melting temperature change curve of the period is substituted into the first... Periodic melting temperature vector For each missing data acquisition moment, the corresponding supplementary melting temperature is calculated to generate the first... Periodic alignment of melting temperature vector Due to the melting temperature Indirectly affects mold temperature The modeling effect cannot be displayed; a time-series prediction network is used to analyze the mold temperature. To complete the feature extraction, the temporal prediction network includes a feature construction layer and a long short-term memory layer. The feature construction layer extracts the first... The melting temperature or supplementary melting temperature at each acquisition moment in the cycle is multiplied by the correlation between the melting temperature and the mold temperature to obtain the [number]th [period]. The weighted melting temperature at each acquisition moment in the cycle will be the first... Periodic mold temperature vector The mask value for each missing acquisition moment is marked as 0, and the mask value for each non-missing acquisition moment is marked as 1. Periodic mold temperature vector Interpolating 0 as the mold temperature at each missing acquisition moment, the first... The weighted melting temperature, mask value, and mold temperature at each acquisition time in the cycle are concatenated column by column to form the first... The mold feature vector at each acquisition time of the period is input into the long short-term memory layer. If the first... If the mask value at each acquisition moment in the cycle is 1, then jump to the next acquisition moment; if the mask value at each acquisition moment is 1, then jump to the next acquisition moment. The mask value is 0 at each acquisition time of the period, and the first... The mold feature vector at each acquisition time of the period and the first The mold temperature or supplementary mold temperature from the previous acquisition time in the cycle is concatenated in columns and processed collaboratively by the input gate, forget gate, and output gate to obtain the [number]th [cycle]. The cell state at each acquisition time point in the cycle is obtained by mapping the data using linear modulation and the ReLU function. The temperature of the mold is replenished at each acquisition time in the cycle, and the cycle continues until the temperature of the mold is replenished. Periodic At the [number] acquisition time, the [number]th data point is generated. Periodic alignment of the mold temperature vector ;

[0053] For the second interpolation subchain, due to the motor current... Directly affects the liquid rise rate And the rate of liquid rise With holding pressure and hydraulic pressure In a strongly correlated, low-pressure casting machine, the liquid lifting process is sequentially executed by a three-phase motor, hydraulic pump, hydraulic cylinder, and lift pipe. The output power of the three-phase motor... equal times the rated operating voltage Motor current With power factor The product of, and the output power Determines the output flow rate of the hydraulic pump's molten metal. Output flow Equal to 1000 times the output power Divide by the pump inlet and outlet pressure difference With pump efficiency The product of, and the liquid rise rate Equal to output flow Divide by the area of ​​the hydraulic cylinder The liquid-liquid mapping model is integrated and constructed as follows:

[0054] ,

[0055] For the periodic acceleration vector For each missing data collection moment, the first... Periodic current vector Substituting the motor current at the corresponding moment into the liquid lifting mapping model, we obtain the first... The replenishment fluid rate at each missing acquisition moment in the cycle constitutes the first... Periodic alignment acceleration vector When molten metal flows in the riser pipe, the holding pressure is... To overcome flow resistance, the Bernoulli equation is used, based on fluid mechanics, to establish the associated holding pressure. With the liquid rise rate The pressure holding mapping model is as follows:

[0056] ,

[0057] in, The density of the molten metal, Atmospheric pressure, , and These are the resistance coefficient, length, and inner diameter of the riser pipe, respectively, due to the pressure holding pressure. With the liquid rise rate Collected at the same frequency, so the first Periodic holding pressure vector Missing data collection time and the first periodic acceleration vector The same, will the first Periodic alignment acceleration vector Substituting the replenishment fluid rise rate into the pressure holding mapping model sequentially, we obtain the first... Periodic holding pressure vector The supplementary holding pressure at each missing acquisition moment constitutes the first... Periodic alignment holding pressure vector Due to the hydraulic pressure inside the hydraulic cylinder The frictional force of the hydraulic cylinder and the reaction force of the molten metal need to be kept in balance to maintain the system force balance. The frictional force of the hydraulic cylinder and the reaction force of the molten metal are related to the lifting speed. There is a potential correlation. Using the temporal prediction network applied in the first interpolation subchain, the [predicted] sequence is extracted. Periodic alignment acceleration vector The liquid rise rate or supplemental liquid rise rate at each acquisition time is multiplied by the liquid rise rate. With hydraulic pressure The degree of correlation between them is obtained. The weighted liquid rise rate at each acquisition moment in the cycle will be the first... Periodic hydraulic vector The mask values ​​for missing and non-missing acquisition times are marked as 0 and 1, respectively, in the first... Periodic hydraulic vector Interpolating 0 as the hydraulic pressure at each missing acquisition moment, the first is constructed in the same way as generating the mold feature vector. The hydraulic feature vector at each acquisition time of the cycle is input into the long short-term memory layer, based on the first... The mask value decision at each acquisition time point in the cycle directly jumps to the next acquisition time point or maps the output. The replenished hydraulic pressure is adjusted at each acquisition time point in the cycle, and the cycle continues until the pressure is replenished. Periodic At the [number] acquisition time, the [number]th data point is generated. Periodic alignment of hydraulic vectors It is worth noting that the rated operating voltage V, power factor Pump efficiency .

[0058] The monitoring and analysis module uses the physical properties of six parameters to determine the melting temperature. and mold temperature Divide into temperature groups and set holding pressure and hydraulic pressure Divide into pressure groups and adjust the liquid lifting rate. and motor current Divide into dynamic groups and obtain the first... periodic alignment parameter matrix , will the Periodic alignment of melting temperature vector and Alignment Mold Temperature Vector Concatenate the rows into a temperature submatrix, and then concatenate the rows into the temperature submatrix. Periodic alignment holding pressure vector and aligned hydraulic vector Concatenate the rows into a pressure submatrix, and then concatenate the rows into the pressure submatrix. Periodic alignment acceleration vector and current vector Row-wise concatenation creates dynamic submatrices. For the three submatrices, temporal extraction is performed using one-dimensional cosine convolution. This one-dimensional cosine convolution uses a cosine kernel instead of a linear convolution kernel to obtain the first submatrix. The system comprises a temperature sub-time feature matrix, a pressure sub-time feature matrix, and a dynamic sub-time feature matrix. The cosine kernel, compared to the traditional linear convolution kernel, uses a cosine curve kernel function, eliminating the need to learn kernel weights and significantly improving computational efficiency. Furthermore, the cosine kernel's responses in the high-frequency and low-frequency bands correspond to abrupt changes and steady-state shifts in parameter time series, respectively, demonstrating strong feature interpretability. Attention mechanisms are applied to the three sub-matrices to obtain the first... The temperature subspace eigenvector, pressure subspace eigenvector, and dynamic subspace eigenvector of the period will be the first... The temperature sub-time characteristic matrix, pressure sub-time characteristic matrix, and dynamic sub-time characteristic matrix of the period are flattened and concatenated column-wise with the corresponding subspace characteristic vectors to obtain the th... The system includes temperature sub-feature vectors, pressure sub-feature vectors, and dynamic sub-feature vectors. The kernel size and stride of the cosine one-dimensional convolution are 3 and 1, respectively. The padding method is Same, and the attention mechanism adopts a single-head attention structure. The scaling factor is the reciprocal of the square root of the sub-feature vector dimension.

[0059] like Figure 4As shown, further, the monitoring and analysis module defines a state space and an action space for each sub-matrix. The state space includes the sub-feature vector sample set of the historical operating conditions corresponding to the sub-matrix, the labeled no-fault operating condition sample labels, the fault operating condition sample labels, and the... The sub-feature vectors of the sub-matrix in the periodicity, the fault-free condition samples are standard samples of the parameters corresponding to the sub-matrix collected when the equipment is running normally, and the fault condition samples are abnormal samples of the parameters corresponding to the sub-matrix collected when the equipment experiences a corresponding type of fault. The action space is the adjustable range of the SNN judgment threshold corresponding to the sub-matrix, specifically including the correction coefficient range of the impulse variation threshold of the sub-matrix. Each action corresponds to a set of threshold correction coefficients, used to dynamically adjust the initial SNN judgment threshold of the sub-matrix, with the optimization goal of maximizing the fault screening accuracy of the sub-matrix. The reward function includes a positive reward term and a negative penalty term. When the... When the SNN initial screening result of the submatrix in a given period matches the actual operating condition labeling result, a positive reward is given. The reward value is positively correlated with the initial screening accuracy of the submatrix. The initial screening accuracy is the proportion of samples whose judgment results match the actual operating conditions in the historical iterations of the submatrix to the total number of test samples. The negative penalty includes two types of core penalties, with the weight of the missed judgment penalty being higher than that of the misjudgment penalty, prioritizing the avoidance of the safety risk of equipment faults not being identified. The first type is the penalty for misjudging a fault-free operating condition. When the submatrix is ​​actually a fault-free operating condition, but the SNN initial screening misjudges it as a fault, a penalty is given based on the square of the misjudgment rate, which is the proportion of fault-free samples that are misjudged as faulty. The second type is the penalty for missed judgment of a faulty operating condition. When the submatrix is ​​actually a faulty operating condition, but the SNN initial screening misses it... When a sample is classified as fault-free, a penalty is applied based on the cube of the false negative rate, where the false negative rate is the proportion of faulty samples that were mistakenly classified as fault-free. First, multiple rounds of offline iterative training are performed using historical labeled working condition samples from the sub-matrix. In each iteration, a threshold correction coefficient is selected based on the action space, an initial screening using an SNN is performed, the corresponding reward value is calculated, the Q-value table is updated, and the process converges to the optimal policy, yielding the impulse mutation threshold for the sub-matrix. The sub-feature vectors of each sub-matrix in each period, along with the actual feedback results after fault determination, are input into the reinforcement learning framework. The Q-value table for each sub-matrix is ​​then updated incrementally online, dynamically correcting the impulse mutation threshold for the next period. Q-learning is a mature reinforcement learning technique; the update method is not detailed here. The correction coefficient range for the impulse mutation threshold is... The ratios of positive reward weight, missed judgment penalty weight, and misjudgment penalty weight are 2:3:1, respectively. The discount factor, learning rate, and training iterations involved in Q-learning are 0.95, 0.2, and 10000, respectively.

[0060] Furthermore, the monitoring and analysis module evaluates the data based on the initial screening by the SNN optimized through Q-learning, and then... Half the difference between the maximum and minimum sub-eigenvalues ​​of the sub-eigenvectors of each sub-matrix in a period is used as the pulse excitation threshold for each sub-matrix. When each sub-eigenvalue in the sub-eigenvectors of each sub-matrix is ​​greater than the corresponding pulse excitation threshold, a pulse is marked as 1 to indicate an excitation pulse; when each sub-eigenvalue in the sub-eigenvectors of each sub-matrix is ​​less than or equal to the corresponding pulse excitation threshold, a pulse is marked as 0. A pulse sequence with the same dimension as the sub-eigenvectors of each sub-matrix is ​​generated. The pulse variation coefficient of each sub-matrix is ​​obtained based on the ratio of the standard deviation to the mean of the pulse sequence of each sub-matrix. Pulse sequences with a pulse variation coefficient greater than the first... The first pulse variation threshold of the period The submatrix of the period is determined to be faulty, and the pulse variation coefficient is less than or equal to the first... The first pulse variation threshold of the period The periodic submatrix is ​​determined to be fault-free.

[0061] Furthermore, the monitoring and analysis module is based on the first The fault sub-matrix of the cycle is either the temperature sub-time characteristic matrix, the pressure sub-time characteristic matrix, or the dynamic sub-time characteristic matrix. The sub-feature vectors of the periodic fault sub-matrix are input to the corresponding temperature branch mapping layer, pressure branch mapping layer, or dynamic branch mapping layer. All three branch mapping layers employ a nested structure of two fully connected layers. The first fully connected layer expands the dimension of the feature vectors of each sub-matrix through linear modulation and enhances nonlinear expressive power using the ReLU function. The second fully connected layer adjusts the output of the first fully connected layer to the same dimension as the total number of branch fault types corresponding to the sub-matrix and maps it to the corresponding branch fault probability vector using the Softmax function. The output branch fault probability vector shows the branch fault types corresponding to branch fault probabilities greater than a probability threshold. Branch faults include temperature faults, pressure faults, and dynamic faults, and each branch fault includes multiple branch fault types. Based on the... The fault submatrix of the period determines the corresponding grouping and extracts the group template vector corresponding to the group. The group template vector corresponding to the group is then calculated and compared with the first... The normalized Euclidean distance of the sub-feature vectors of the periodic fault submatrix is ​​used to determine the corresponding fault level based on the fault level range to which the normalized Euclidean distance belongs. The lower the fault level, the more minor the fault. The fault level range is preset, and the groups include temperature group, pressure group, and dynamic group. In this embodiment, when the normalized Euclidean distance is greater than 0 and less than or equal to 0.3, it is determined as a level 1 fault. When the normalized Euclidean distance is greater than 0.7 and less than or equal to 1, it is determined as a level 3 fault. All other cases are determined as level 2 faults. The number of neurons in the first fully connected layer is 128.

[0062] Specifically, in this embodiment, the temporal prediction network, the fully connected network in the branch mapping layer, and the spiking neural network (SNN) all require pre-training. The temporal prediction network uses a long short-term memory network (LSTM) as its core architecture. The dataset is collected from historical data collected over 1500 operating cycles of the target low-pressure casting machine, including 1200 normal operating cycles and 300 non-steady-state operating cycles with parameter fluctuations. The dataset covers the entire operating range of melting temperature, mold temperature, liquid lifting speed, and hydraulic pressure, and is divided into training, validation, and test sets in an 8:1:1 ratio. During the dataset preprocessing stage, Z-Sco is performed on all parameter vectors. For re-standardization, according to the aforementioned rules of this scheme, mask values ​​of 1 and 0 are set for non-missing acquisition times and missing acquisition times, respectively, to construct time-series training samples with mask labels. The Adam adaptive optimization algorithm is used to complete the training, with the mask mean square error (MSE) as the loss function. The loss is calculated only for the imputation results of missing acquisition times with a mask value of 0. The initial learning rate is set to 0.001, and the learning rate decays by 0.9 times every 10 training epochs. The training batch size is 32, and the maximum training epochs are 200. An early stopping strategy is set, and training is terminated when the validation set loss does not decrease for 15 consecutive training epochs. The ReLU function is used as the network activation function.

[0063] The fully connected network in the branch mapping layer for fault type identification constructs independent networks with the same architecture for three sub-matrices: temperature, pressure, and dynamic. The dataset includes: sub-feature vector samples with fault type annotations collected during the historical operation of the target low-pressure casting machine, and sub-feature vector samples covering all types of faults generated through fault injection simulation. In total, it contains 4000 sets of normal operating condition samples, 2500 sets of temperature-related fault samples, 3500 sets of pressure-related fault samples, and 2000 sets of dynamic fault samples. These are divided into training, validation, and test sets in a 7:2:1 ratio. The feature vectors are normalized to match the normalized Euclidean distance calculation logic for fault level determination. The Adam adaptive optimization algorithm is used for training, with multi-class cross-entropy as the loss function and L2 regularization added to avoid overfitting. The initial learning rate is set to 0.0005, the training batch size is 64, and the maximum number of training epochs is 150. An early stopping strategy is set, and training is terminated when the classification accuracy on the validation set does not improve for 10 consecutive training epochs. The activation function of the first fully connected layer is the ReLU function, and the output of the second fully connected layer is mapped to the fault probability vector using the Softmax function.

[0064] The Spiking Neural Network (SNN) used for initial fault screening employs a Q-learning reinforcement learning algorithm as its core optimization framework. Training and optimization are completed using the SNN's spike sequence generation and fault determination logic, divided into two stages: offline pre-training and online incremental adaptation. In the offline pre-training stage, the dataset was collected from 2100 operating cycles of a target low-pressure casting machine, including 1500 sets of fault-free operating condition samples and 600 sets of fault-annotated operating condition samples, covering all fault scenarios across three sub-matrices. This data forms an experience replay pool of 1000 samples. Samples are stored separately according to sub-matrix type, and independent reinforcement learning optimization branches are constructed for each of the three sub-matrix types. An ε-greedy strategy is used to balance exploration and utilization. The Q-value table is iteratively updated based on the Bellman equation, with a discount factor of 0.95 and a learning rate of 0.2. The initial exploration rate is 0.9, and the exploration rate decays by 0.05 times every 1000 training rounds, eventually converging to 0.1. The maximum number of iterations for offline pre-training is 10,000 rounds. The reward function is constructed according to the rules mentioned above in this scheme, with the accuracy of fault screening as the positive reward and the false positive rate and false negative rate of fault-free conditions as the negative penalty. The penalty weight for false negatives is 3 times that for false negatives, prioritizing the fault detection rate. In the online incremental adaptation phase, each cycle of real-time equipment operation is used as an update unit. The sub-feature vectors, fault screening results, and actual operating condition feedback data in each cycle are added to the experience replay pool. The Q-value table is incrementally updated with a learning rate of 0.05. Simultaneously, the pulse mutation threshold corresponding to each sub-matrix is ​​dynamically optimized to ensure that the screening threshold continuously adapts to the real-time operating status and long-term operating condition drift characteristics of the equipment.

[0065] This invention discloses a remote fault monitoring system for low-pressure casting machines, including a data acquisition and processing module, a monitoring and analysis module, and an alarm module. The data acquisition and processing module collects melting temperature, mold temperature, holding pressure, hydraulic pressure, liquid lifting speed, and motor current at differentiated sampling frequencies. It combines a 5-level progressive decomposition based on the db4 wavelet basis with soft thresholding based on the Donohue threshold to achieve parameter noise reduction, adaptively balancing noise suppression and parameter detail preservation. Through grey relational analysis of historical co-occurring parameters, and by combining process flow, relational thresholds, and stability thresholds to screen and construct parameter interpolation chains, it employs cubic spline fitting, a liquid lifting mapping model, and a time-series prediction network to complete missing parameters, achieving parameter accuracy that aligns with process logic. The system performs time-series alignment; the monitoring and analysis module splits the alignment parameter matrix into temperature, pressure, and dynamic sub-matrices, extracts temporal features through cosine one-dimensional convolution, extracts spatial features through an attention mechanism and concatenates them to generate sub-feature vectors, and uses reinforcement learning to dynamically optimize the SNN judgment threshold and quickly judge faults based on the SNN initial screening, improving the accuracy and adaptability of the initial screening. The fault type is determined by the branch mapping layer corresponding to the sub-matrix, and the fault level is determined based on the normalized Euclidean distance between the sub-feature vector and the group template vector, improving the efficiency and accuracy of fault judgment; the alarm module executes graded alarms according to the fault level to ensure timely transmission of fault information and ensure the safe operation and stable production of the low-pressure casting machine.

[0066] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A remote fault monitoring system for low-pressure casting machines, characterized in that, It includes a data acquisition and processing module, a monitoring and analysis module, and an alarm module; The data acquisition and processing module collects melting temperature, mold temperature, holding pressure, hydraulic pressure, liquid lifting speed and motor current at different sampling frequencies. In each cycle, it organizes them into a parameter vector set and performs wavelet transform noise reduction. Based on the parameter interpolation chain, it selects a model for each parameter in sequence for interpolation, integrates them to obtain the alignment parameter matrix of each cycle and marks the equipment number. The parameter interpolation chain is obtained by performing grey relational analysis and multidimensional screening on the parameters at the co-occurring time in the historical cycles. The monitoring and analysis module splits the alignment parameter matrix of each cycle into three sub-matrices. It extracts the temporal and spatial features of each sub-matrix through cosine one-dimensional convolution and attention mechanism and integrates them into sub-feature vectors. It uses an SNN optimized by Q-learning to screen and judge whether each sub-matrix is ​​faulty. The sub-feature vector of the faulty sub-matrix is ​​used to determine the fault type through the corresponding branch mapping layer. The fault level is determined based on the normalized Euclidean distance between the sub-feature vector of the faulty sub-matrix and the group template vector of the corresponding group. Finally, it generates a fault report by combining the equipment number. The alarm module triggers tiered alarms based on fault reports.

2. The remote fault monitoring system for low-pressure casting machines as described in claim 1, characterized in that, The acquisition and processing module extracts six parameters from the parameter vector set of historical cycles at all co-occurring moments, splices and integrates them, and performs standardization to obtain a standard collinear parameter matrix. Based on grey relational analysis, it calculates the correlation between parameters and constructs an undirected parameter graph. Referring to the process flow, it sets directional constraints to convert undirected edges into directed edges. After pruning the remaining undirected edges, it filters directed edges based on correlation and stability thresholds and constructs parameter interpolation chains.

3. The remote fault monitoring system for low-pressure casting machines as described in claim 1, characterized in that, The acquisition and processing module uses wavelet transform based on the db4 wavelet basis for noise reduction. It performs a 5-level progressive decomposition on the parameter vector composed of each parameter in the parameter vector set to generate a low-frequency coefficient vector of the 5th level and a high-frequency coefficient vector of the 5th level. It calculates the absolute median difference of the high-frequency coefficient vector of the 1st level and divides it by 0.6745 to obtain the noise standard deviation. It multiplies the square root of the natural logarithm of the parameter vector dimension by the noise standard deviation to obtain the Donogh threshold. It performs soft thresholding on the high-frequency coefficient vector to obtain a 5-level denoised high-frequency coefficient vector.

4. The remote fault monitoring system for low-pressure casting machines as described in claim 1, characterized in that, The monitoring and analysis module uses half the difference between the maximum and minimum sub-eigenvalues ​​of the sub-eigenvectors of each sub-matrix as the pulse excitation threshold. It compares each sub-eigenvalue in the sub-eigenvectors of each sub-matrix with the pulse excitation threshold to set a pulse marker, generates a pulse sequence, calculates the ratio of the standard deviation to the mean of the pulse sequence to obtain the pulse variation coefficient, and compares the pulse variation coefficient with the pulse variation threshold of the current period to determine whether the sub-matrix is ​​faulty. The pulse variation threshold of the current period is obtained by combining the sub-eigenvectors of each sub-matrix in the previous period with the actual judgment results and Q-learning optimization.

5. The remote fault monitoring system for low-pressure casting machines as described in claim 1, characterized in that, The monitoring and analysis module concatenates the aligned melt temperature vector with the aligned mold temperature vector, the aligned holding pressure vector with the aligned hydraulic pressure vector, and the aligned acceleration vector with the aligned current vector in the alignment parameter matrix according to physical properties, into temperature sub-matrices, pressure sub-matrices, and dynamic sub-matrices, respectively. It then uses cosine one-dimensional convolution to extract the sub-time feature matrix of each sub-matrice and flattens it. Finally, it generates the sub-space feature vector of each sub-matrice through an attention mechanism and concatenates them to obtain the sub-feature vector of each sub-matrice.

6. The remote fault monitoring system for low-pressure casting machines as described in claim 1, characterized in that, The monitoring and analysis module inputs the sub-feature vectors of the fault submatrix into the corresponding branch mapping layer, maps them into branch fault probability vectors through two fully connected layers, outputs the fault type corresponding to the branch fault probability that is greater than the probability threshold, extracts the group template vector of the corresponding group of the fault submatrix and calculates the normalized Euclidean distance with the sub-feature vector, and determines the fault level based on the belonging interval of the normalized Euclidean distance.

7. The remote fault monitoring system for low-pressure casting machines as described in claim 2, characterized in that, The acquisition and processing module refers to the first interpolation subchain, performs cubic spline fitting on the melt temperature vector, and calculates the supplementary melting temperature at the missing acquisition time to generate an aligned melt temperature vector. The aligned melt temperature is multiplied by the correlation between the melting temperature and the mold temperature to obtain the weighted melting temperature. Different mask values ​​are set for the acquisition time and the missing acquisition time of the mold temperature vector, and the initial mold temperature is filled in at the missing acquisition time. The temporal prediction network is used to combine the weighted melting temperature, mask value and mold temperature to perform temporal supplementation for the missing acquisition time, and the initial mold temperature is replaced with the corresponding supplementary mold temperature to generate an aligned mold temperature vector.

8. The remote fault monitoring system for low-pressure casting machines as described in claim 7, characterized in that, The acquisition and processing module, referencing the second interpolation subchain, constructs a liquid-lifting mapping model, with the liquid-lifting rate equal to 1000. The product of the rated operating voltage, motor current, and power factor is divided by the product of the hydraulic cylinder area, pump inlet / outlet pressure difference, and pump efficiency. The supplementary liquid lifting speed at each missing acquisition moment in the lifting speed vector is calculated sequentially to form an aligned lifting speed vector. A pressure holding mapping model relating the pressure holding pressure and the liquid lifting speed is established using the Bernoulli equation. The pressure holding pressure is equal to a constant coefficient multiplied by the square of the liquid lifting speed plus atmospheric pressure. The supplementary pressure holding pressure at each missing acquisition moment in the pressure holding vector is calculated sequentially to form an aligned pressure holding vector. An aligned hydraulic pressure vector is generated by combining a time-series prediction network with the aligned lifting speed vector in the same way as the aligned mold temperature vector.

9. The remote fault monitoring system for low-pressure casting machines as described in claim 3, characterized in that, The acquisition and processing module uses the db4 wavelet basis to perform inverse transform reconstruction on the low-frequency coefficient vector of the 5th layer and the noise-reduced high-frequency coefficient vector of the 5th layer in each cycle. After upsampling the low-frequency coefficient vector and high-frequency coefficient vector of each layer by a factor of two, they are transposed and convolved with the low-pass filter and the high-pass filter respectively and summed to obtain the low-frequency coefficient vector of the previous layer in each cycle, until the reconstruction reaches the 0th layer, generating the pure parameter vector of each cycle.

10. The remote fault monitoring system for low-pressure casting machines as described in claim 1, characterized in that, The data acquisition and processing module collects the melting temperature and mold temperature according to the sampling frequency, collects the holding pressure, hydraulic pressure and liquid lifting speed at twice the sampling frequency, and collects the motor current at four times the sampling frequency. In each cycle, the data are statistically spliced ​​into melting temperature vector, mold temperature vector, holding pressure vector, hydraulic pressure vector, lifting speed vector and current vector, and combined to construct the parameter vector set for each cycle.

Images