Intelligent control method and system for multiple machines

Abstract

This invention relates to the field of intelligent control technology, and discloses an intelligent control method and system for multi-machine collaboration. The method includes: fusing multi-machine environment and single-machine state data, extracting time-series features to obtain single-machine state deviation values; if the deviation exceeds a preset threshold, extracting the current operating cycle time, retrieving adjacent machines through a machine topology map and embedding their process sequences to obtain associated operating cycle times; calculating the difference between the current cycle time, associated cycle time, and standard cycle time to obtain cycle time loss, extracting the initial feed rate and cooling time accordingly, adjusting them through cross-coupling to obtain target values, and generating a synchronized cluster timing sequence; if the misalignment between this timing sequence and the standard timing sequence exceeds the limit, generating a state correction command to align the timing of the multiple machines and obtain collaborative control results. This method can achieve dynamic perception and adaptive collaboration of multi-machine operating cycle times.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an intelligent control method and system for multi-machine collaboration. Background Technology

[0002] Currently, in the fields of precision manufacturing and automated production, multi-machine collaborative operation has become the core mode for improving production line efficiency and product consistency. The various processing machines form a tightly coupled production link through material flow and process connection. Fluctuations in the operating cycle of any machine will be transmitted step by step along the material flow direction, eventually causing the entire cluster to lose synchronization and reduce capacity.

[0003] In existing technologies, multi-machine collaborative control often employs centralized planning methods based on fixed cycle times or simple master-slave synchronization strategies executed by the industrial control system. These methods primarily focus on whether each machine completes its designated actions according to a preset time sequence, correcting for deviations between the actual and standard cycle times through periodic polling. However, when fluctuations in workshop ambient temperature cause changes in coolant supply pressure, or when tool wear on a single machine leads to a passive reduction in feed rate, the single cycle time comparison logic cannot identify the coupled impact of environmental factors and individual machine degradation on the cluster's timing. The industrial control system can only passively intervene after the deviation accumulates to trigger an alarm threshold, lacking early detection of out-of-sync trends and comprehensive consideration of the impact on adjacent machines, resulting in frequent production line stoppages and manual intervention.

[0004] In summary, existing technologies lack the ability to comprehensively perceive changes in the environment and the status of individual machines, making it difficult to dynamically predict and adaptively coordinate the operating cycle of multiple machines under complex operating conditions, and resulting in difficulties in maintaining the consistency of cluster timing. Summary of the Invention

[0005] This invention provides an intelligent control method and system for multi-machine collaboration, so as to realize dynamic prediction and adaptive collaborative control of the operating cycle of multiple machines.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent control method for multi-machine collaboration, comprising:

[0007] By integrating environmental change data and individual machine status data from multiple machines on an industrial production line, global perception information is obtained, and the temporal features of the global perception information are extracted to obtain the individual machine status deviation value.

[0008] If the deviation value of the single machine status is greater than the preset deviation threshold, the current running cycle is extracted, the adjacent machines are obtained by searching the machine topology map, and the process flow sequence of the adjacent machines is processed by sequence embedding to obtain the associated running cycle.

[0009] The difference between the current running cycle, the associated running cycle, and the preset standard running cycle is calculated to obtain the cycle step loss amount;

[0010] The initial feed rate and initial cooldown time are extracted based on the cycle step loss. The initial feed rate and initial cooldown time are adjusted by cross-coupling to obtain the target feed rate and target cooldown time. The synchronized adjusted cluster timing is generated based on the target feed rate and target cooldown time.

[0011] If the misalignment between the cluster timing and the preset standard timing is greater than the preset misalignment threshold, a state correction command is generated to align the timing of the multiple machines and obtain the collaborative control result of the multiple machines.

[0012] In one optional implementation, the fusion of environmental change data and individual machine status data from multiple machines on the industrial production line to obtain global perception information includes:

[0013] The environmental change data and the single-machine state data are subjected to time offset compensation processing to obtain an aligned environment sequence and an aligned state sequence;

[0014] The alignment environment sequence and the alignment state sequence are concatenated to obtain an initial fused dataset. The initial fused dataset is then filled in to obtain a complete fused dataset.

[0015] The complete fusion dataset is processed to construct a view, resulting in a perceptual fusion view. Global perceptual information is then obtained from the perceptual fusion view.

[0016] In one optional implementation, the step of extracting the temporal features of the global perception information to obtain the single-machine state deviation value includes:

[0017] The global perception information is sequence-segmented using a long short-term memory network to obtain initial temporal features;

[0018] The initial time series features are nonlinearly mapped to obtain deep time series features;

[0019] If the feature dimension of the deep temporal feature is greater than a preset dimension threshold, then principal component extraction is performed on the deep temporal feature to obtain the standard temporal feature;

[0020] The state residual is calculated by comparing the standard timing characteristics with the pre-established deviation benchmark to obtain the single-machine state deviation value.

[0021] In one optional implementation, if the deviation value of the single machine state is greater than a preset deviation threshold, then extracting the current operating cycle and retrieving adjacent machines through the machine topology map includes:

[0022] If the deviation value of the single machine state is greater than the preset deviation threshold, then the beat timestamp and the running action cycle corresponding to the single machine state deviation value are extracted, and the beat timestamp and the running action cycle are combined and encapsulated to obtain the current running beat.

[0023] The adjacent machines are obtained by searching the adjacent nodes of the current machine through the machine topology graph.

[0024] In one optional implementation, the step of performing sequence embedding processing on the process flow sequences of adjacent machines to obtain associated operating cycles includes:

[0025] Obtain the process flow sequence of adjacent machines;

[0026] The process flow sequence is subjected to sequence embedding processing to obtain a beat feature vector. The cosine similarity between the current running beat and the beat feature vector is calculated to obtain the association matching degree.

[0027] If the correlation matching degree is greater than the preset matching threshold, then the associated running beat is extracted from the beat feature vector.

[0028] In one optional implementation, the step of calculating the difference between the current running cycle, the associated running cycle, and the preset standard running cycle to obtain the cycle time loss includes:

[0029] The current running beat, the associated running beat, and the preset standard running beat are aligned on the time axis to obtain a time offset sequence;

[0030] Extract process flow delays from the time offset sequence, and construct a cycle time difference matrix based on the process flow delays;

[0031] The beat difference matrix is ​​subjected to feature extraction to obtain the out-of-step feature vector. The out-of-step feature vector is then mapped to nodes in a preset multi-machine collaborative graph to obtain the beat out-of-step amount.

[0032] In one optional implementation, the step of extracting the initial feed rate and initial cooldown time based on the cycle time loss, adjusting the initial feed rate and initial cooldown time through cross-coupling to obtain the target feed rate and target cooldown time, and generating a synchronized adjusted cluster timing sequence based on the target feed rate and target cooldown time includes:

[0033] Based on the cycle step loss, extract the initial feed rate of the current machine and the initial cooling time of the adjacent machine;

[0034] The difference coupling calculation is performed between the current operating cycle time of the current machine and the associated operating cycle time of the adjacent machine to obtain the multi-machine synchronization coupling error;

[0035] A cross-coupling control algorithm is adopted to reduce the initial feed rate based on the multi-machine synchronous coupling error to obtain the target feed rate, and to extend the initial cooling time based on the reverse compensation component of the multi-machine synchronous coupling error to obtain the target cooling time.

[0036] A feedback matrix is ​​constructed based on the target feed rate and the target cooling time. The feedback matrix is ​​then encapsulated using a protocol to obtain control instructions. These control instructions are then sent to the programmable logic controller to obtain the synchronized cluster timing.

[0037] In one optional implementation, if the misalignment between the cluster timing and the preset standard timing is greater than a preset misalignment threshold, a state correction command is generated to perform timing alignment on the multiple machines to obtain a multi-machine collaborative control result, including:

[0038] The synchronized cluster timing sequence is compared with the preset standard timing sequence to obtain the timing deviation sequence;

[0039] The time-series deviation sequence is reduced in dimensionality to obtain the core misalignment amount. If the core misalignment amount is greater than a preset misalignment threshold, the computer station state compensation matrix is ​​calculated based on the core misalignment amount.

[0040] A command sequence is constructed based on the machine state compensation matrix, and the command sequence is converted to a protocol to obtain a state correction command.

[0041] The response delay sequence of each machine is extracted according to the state correction command. An alignment compensation matrix is ​​calculated based on the maximum delay value in the response delay sequence. The alignment compensation matrix is ​​used to modulate the underlying drive signal to obtain a modulated drive signal. The operation actions of multiple machines are time-aligned using the modulated drive signal to obtain the multi-machine collaborative control result.

[0042] Secondly, the present invention provides a multi-machine collaborative intelligent control system, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-machine collaborative intelligent control method as described above.

[0043] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent control method for multi-machine collaboration as described above.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) This invention integrates environmental change data from multiple machines with single-machine status data, performs time-series offset compensation and splicing filling on heterogeneous sensor data, constructs a perception fusion view, obtains global perception information, and then extracts time-series features to obtain single-machine status deviation values. This series of operations effectively breaks down the data silos between workshop environmental information and single-machine operation information. By using a long short-term memory network to perform deep feature extraction and principal component dimensionality reduction on multi-source time-series data, it achieves accurate quantitative perception of single-machine abnormal states, solving the problem in existing technologies where industrial control systems cannot perceive the coupled effects of environmental factors and single-machine degradation.

[0046] (2) When the deviation value of a single machine exceeds a preset threshold, this invention extracts the current operating cycle and retrieves adjacent machines through the machine topology graph. The process flow sequence of the adjacent machines is processed by sequence embedding to obtain the associated operating cycle. Then, the difference between the current cycle, the associated cycle, and the standard cycle is calculated to obtain the cycle step loss. By using the directed edge weight screening and cosine similarity matching of the machine topology graph, it can accurately identify upstream and downstream machines with real cycle step linkage relationships, avoiding the misadjustment problem caused by blindly associating all adjacent machines, and realizing the accurate location of the root cause of the step loss and the quantitative assessment of the degree of step loss.

[0047] (3) This invention extracts the initial feed rate and initial cooling time based on the step loss amount, and uses a cross-coupling control algorithm to coordinate the adjustment of the two to generate a synchronized cluster timing sequence. The cross-coupling control simultaneously distributes the synchronization coupling error of multiple machines to the local controller of each machine for coordinated correction. Compared with the traditional single-variable PID control, it has better multi-variable coordinated regulation capability, avoids dimensional mismatch and parameter mutation disturbance caused by adjusting two different physical quantity parameters at the same time, so that the regulation process can match the severity of step loss and maintain the smooth transition of cluster timing sequence, significantly improving the adaptive regulation capability of the cluster under complex working conditions.

[0048] (4) When the misalignment between the cluster timing and the standard timing exceeds the misalignment threshold, this invention generates a state correction command for timing alignment. By extracting the maximum delay value from the response delay sequence of each machine tool, the alignment compensation matrix is ​​calculated, and the underlying drive signal is modulated to achieve precise timing alignment of the operation of multiple machines. This closed-loop correction mechanism eliminates the residual misalignment caused by the difference in actuator response, effectively ensuring the timing consistency and collaborative stability of the multi-machine system under all operating conditions, and improving the continuous operation reliability and product quality consistency of the precision manufacturing production line. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the intelligent control method for multi-machine collaboration provided in the first embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Reference Figure 1 The first embodiment of the present invention provides an intelligent control method for multi-machine collaboration, comprising the following steps:

[0052] Step S11: Integrate environmental change data and single machine status data of multiple machines on the industrial production line to obtain global perception information, extract the temporal features of the global perception information, and obtain the single machine status deviation value.

[0053] Step S12: If the deviation value of the single machine status is greater than the preset deviation threshold, the current running cycle is extracted, the adjacent machines are obtained by searching the machine topology map, and the process flow sequence of the adjacent machines is processed by sequence embedding to obtain the associated running cycle.

[0054] Step S13: Calculate the difference between the current running cycle, the associated running cycle, and the preset standard running cycle to obtain the cycle step loss amount;

[0055] Step S14: Extract the initial feed rate and initial cool-down time based on the cycle step loss amount, and adjust the initial feed rate and initial cool-down time through cross-coupling to obtain the target feed rate and target cool-down time. Generate the synchronized adjusted cluster timing based on the target feed rate and target cool-down time.

[0056] Step S15: If the misalignment between the cluster timing and the preset standard timing is greater than the preset misalignment threshold, a state correction command is generated to perform timing alignment on the multiple machines to obtain the multi-machine collaborative control result.

[0057] In step S11, environmental change data from multiple machines and individual machine status data are fused to obtain global perception information. The temporal features of the global perception information are extracted to obtain the individual machine status deviation value.

[0058] In one implementation, the method includes:

[0059] The environmental change data and the single-machine state data are subjected to time offset compensation processing to obtain an aligned environment sequence and an aligned state sequence;

[0060] The alignment environment sequence and the alignment state sequence are concatenated to obtain an initial fused dataset. The initial fused dataset is then filled in to obtain a complete fused dataset.

[0061] The complete fusion dataset is processed to construct a view, resulting in a perceptual fusion view. Global perceptual information is then obtained from the perceptual fusion view.

[0062] In one implementation, the method further includes:

[0063] The global perception information is sequence-segmented using a long short-term memory network to obtain initial temporal features;

[0064] The initial time series features are nonlinearly mapped to obtain deep time series features;

[0065] If the feature dimension of the deep temporal feature is greater than a preset dimension threshold, then principal component extraction is performed on the deep temporal feature to obtain the standard temporal feature;

[0066] The state residual is calculated by comparing the standard timing characteristics with the pre-established deviation benchmark to obtain the single-machine state deviation value.

[0067] The environmental change data are multi-dimensional environmental parameters collected by distributed sensors in the precision manufacturing workshop, including workshop temperature, humidity, and coolant supply pressure. The sampling period for temperature and humidity sensors is 5 seconds, and the sampling period for coolant pressure sensors is 1 second. The single-machine status data includes the spindle speed, tool vibration frequency, and servo feed load current of each processing machine. The spindle speed and servo feed load current are output by the machine's built-in encoder and current transformer with a sampling period of 1 second. The tool vibration frequency is collected by a piezoelectric accelerometer installed on the outer wall of the spindle box with a sampling frequency of 1000 Hz and then summed according to the root mean square value in a 1-second window.

[0068] Due to differences in sampling periods and network transmission delays among different sensors, there are deviations in the timestamps of data arriving at the edge server. This embodiment employs a simple time delay search method based on the Pearson correlation coefficient to compensate for time offset. The environmental change data sequence is shifted along the time axis according to a preset search range (e.g., 0 to 2s, with a search step size of 100ms). The Pearson correlation coefficient between the two sequences is calculated for each shift. The shift amount that maximizes the correlation coefficient is the time offset. The time offset is then deducted from the corresponding environmental change data timestamp, thus completing the time offset compensation and outputting an aligned environment sequence and an aligned state sequence that are strictly aligned in the time dimension.

[0069] The splicing process involves horizontally merging the aligned multidimensional features according to time nodes, combining environmental parameters and single-machine state parameters at the same physical moment into a single multidimensional record. Each row of the initial fused dataset corresponds to a sampling time, and each column corresponds to a monitored variable. In actual production, due to electromagnetic interference and occasional sensor outages, the initial fused dataset may contain missing data at certain sampling times. A nearest neighbor imputation algorithm is introduced to fill in the missing data: using two adjacent valid sampling points before and after the missing time (a total of four neighbors) as references, the imputation value is obtained by averaging the values ​​based on the inverse of the distance. The physical meaning of the distance weight is that sampling points closer to the missing time have higher relevance and greater contribution. If there are fewer than four available neighbors before and after the missing time (e.g., located at the start and end boundaries of the data), the weighting is based on the actual available neighbors. This imputation strategy achieves a root mean square error of less than 3% on at least 50 sets of field-measured data, thus establishing the complete fused dataset.

[0070] The view construction process uses a pre-constructed machine adjacency graph as the skeleton of the perception fusion view. The machine adjacency graph uses each processing machine in the workshop as a node and the material flow path between machines as directed edges. The edge weights are obtained by weighting and fusing the physical distance and the tightness of the connection between processes according to a preset weight ratio of 0.6:0.4, and then performing min-max normalization. The physical distance is measured through the workshop layout map. The tightness of the connection between processes is obtained by normalizing the maximum and minimum times of the occurrence of two adjacent machines in the process planning document. The preset weight ratio of 0.6:0.4 is obtained by performing correlation analysis between the fused edge weights and the actual coupling degree of manually labeled machines under no less than 30 typical working conditions of the same model, and taking the weight ratio with the maximum correlation coefficient. The correlation coefficient is not less than 0.82. The multidimensional features of each machine in the complete fused dataset are averaged along the time dimension within the current sampling window (e.g., the past 60 seconds) and written to the corresponding node. Then, the features of adjacent nodes are weighted and averaged using edge weights to obtain a perceptual fusion view that integrates spatial topological relationships and temporal statistical features. The global perceptual information is the structured data structure composed of all node features and edge weights within the current sampling window of the perceptual fusion view, serving as the unified input for subsequent temporal feature extraction.

[0071] Next, a Long Short-Term Memory (LSTM) network is used to segment the global perception information into sequences to obtain initial temporal features. Specifically, the LSTM network is a recurrent neural network suitable for processing temporal sequences. To convert the graph-structured perception fusion view into a regular sequence input acceptable to the LSTM, this embodiment uses a node feature serialization method that unfolds according to the machine number: for each machine node in the perception fusion view, the multidimensional features of the node at each sampling time within a preset sliding window (window length 60s, sliding step size 10s) are concatenated temporally to obtain the node temporal segment of that machine; then, the node temporal segments of each machine are spliced ​​together in ascending order of machine number to form a unified sequence segment. Each machine corresponds to the feature vector of 6 consecutive sampling times (i.e., 6 sampling points with a 10s step size within a 60s window). Therefore, a total of 6N×M unified sequence segments are spliced ​​from N machines as LSTM input. In this embodiment, taking 2 machines as an example, the sequence segment dimension is 12×M. Where M represents the feature dimension of each node in the perception fusion view (M=8 in this embodiment), M=8 includes 6 sensor readings (temperature, humidity, coolant pressure, spindle speed, root mean square value of tool vibration, servo load current) and 2 spatial topology derived features. The 2 spatial topology derived features are: the node degree normalized value, which is the number of directly adjacent machines in the machine adjacency graph divided by the maximum node degree in the workshop; and the neighborhood edge weight mean, which is the arithmetic mean of the edge weights between the machine and all its adjacent machines, with a value ranging from 0 to 1. The hidden layer dimension of the LSTM is set to 128, and the number of layers is 1. The LSTM unit uses three gating mechanisms—forget gate, input gate, and output gate—to filter and update the temporal information. The hidden layer state vector corresponding to the last time step is the output, with a dimension of 128, which serves as the initial temporal feature.

[0072] The pre-training process of the LSTM involves collecting at least 5000 sets of historical perception fusion view sequence segments of the target model processing machine under different operating conditions (light load, heavy load, automatic start-stop switching, and external temperature disturbance) in an experimental workshop. The beat stability level (values ​​from 1 to 5) is used as the supervision label. The level scoring details are determined by the on-site engineer according to preset objective scoring rules: the standard deviation σ (in seconds) of the beat time deviation between adjacent machines within the corresponding time window of each sequence segment is collected, and assigned values ​​as follows: σ < 0.5s is level 5 (extremely stable), 0.5s ≤ σ < 1.0s is level 4 (stable), 1.0s ≤ σ < 2.0s is level 3 (basically stable), 2.0s ≤ σ < 4.0s is level 2 (slightly out of step), and σ ≥ 4.0s is level 1 (severely out of step). To eliminate subjective differences among engineers, the score is calculated by averaging the scores given independently by two engineers. When the difference between the two scores is greater than 1, the score is submitted to a third engineer for review. Mean squared error is used as the loss function, Adam optimizer is used, the initial learning rate is 0.001, the number of iterations is no less than 1000, and 5-fold cross-validation is used to prevent overfitting. When the validation set loss no longer decreases for 10 consecutive rounds, an early stopping mechanism is triggered. Finally, the network parameters with fixed weights are saved as the optimal parameters of the model used in this embodiment.

[0073] The nonlinear mapping is achieved through a single fully connected layer cascaded after the LSTM backbone. This fully connected layer processes the 128-dimensional input using the ReLU activation function and outputs 128-dimensional deep temporal features. The ReLU activation function directly sets inputs less than 0 to zero and maintains a linear output for inputs greater than 0, thereby introducing nonlinear expressive power. This allows the deep temporal features to better represent the higher-order coupling relationships in the multi-machine operation process compared to the initial temporal features. The weights of the fully connected layer are jointly trained with the LSTM backbone, and the weight parameters are updated synchronously during the LSTM training phase. The loss gradient updates the parameters of both networks simultaneously through backpropagation.

[0074] The preset dimensionality threshold of 64 is determined based on the real-time requirements of subsequent Euclidean distance calculation and engineering constraints on high-dimensional similarity discrimination capability. Specifically, at least 50 sets of standard time-series feature samples were collected in an experimental workshop for the same model of machine. The Euclidean distance calculation time and similarity discrimination accuracy were tested for dimensions of 16, 32, 64, 128, and 256. When the dimension is greater than 64, the excessive number of squared terms in the Euclidean distance calculation leads to a calculation delay exceeding the 200ms real-time boundary. Furthermore, the curse of dimensionality in high-dimensional space causes the similarity discrimination accuracy to drop by more than 3%. Therefore, 64 is set as the preset dimensionality threshold. In this embodiment, the deep time-series feature dimension is 128, which is greater than the preset dimensionality threshold of 64, triggering dimensionality reduction logic. The principal component extraction uses Principal Component Analysis (PCA) algorithm. PCA is an unsupervised linear dimensionality reduction method that transforms potentially correlated high-dimensional variables into linearly uncorrelated low-dimensional variables by orthogonally transforming the original feature space, and retaining the top principal components with the largest variance.

[0075] In this embodiment, the 128-dimensional deep temporal features are reduced to a dimension of 64. Specifically, the covariance matrix (128×128 dimension) of the deep temporal features is first calculated over the entire training set. Then, eigenvalue decomposition is performed on this covariance matrix, and the eigenvectors corresponding to the top 64 largest eigenvalues ​​are used to construct a projection matrix (128×64 dimension). The real-time acquired 128-dimensional deep temporal features are multiplied by the projection matrix to obtain the 64-dimensional standard temporal features. The projection matrix is ​​calculated and stored once during initialization, requiring only one matrix multiplication during runtime, resulting in minimal computational overhead. The cumulative variance contribution rate retained after dimensionality reduction is greater than 95%, ensuring that information loss is within an acceptable range.

[0076] The deviation reference is pre-established. For each machine in the workshop, a 64-dimensional deviation reference vector is stored. The method of acquisition is to run the machine continuously for no less than 24 hours under standard processing technology and standard environment (temperature 25°C±2°C, humidity 50%±5%RH, no external disturbance) at the end of the machine commissioning period. The average value of all collected standard time-series features is taken according to the time dimension. The resulting 64-dimensional average vector is used as the deviation reference vector of the machine. The deviation reference vector of each machine in the same workshop is established and stored once after production is started. It is recalibrated by maintenance personnel every six months during the downtime maintenance period. The state residual calculation uses Euclidean distance, which involves finding the absolute distance between the current 64-dimensional standard time-series features and the corresponding machine tool's deviation from the baseline vector in a multi-dimensional space: the squared difference of each component is calculated, and the square root of the sum of the squared differences across all 64 dimensions is taken. The result is the single-machine state deviation value, which is a dimensionless scalar; a larger value indicates a more severe deviation of the current machine tool's operating state from the normal baseline. In this embodiment, taking CNC machine tool number 5 as an example, the calculated single-machine state deviation value is 15.

[0077] In step S12, if the deviation value of the single machine state is greater than a preset deviation threshold, the current operating cycle is extracted, adjacent machines are retrieved through the machine topology map, and the process flow sequences of the adjacent machines are subjected to sequence embedding processing to obtain the associated operating cycle. This includes:

[0078] If the deviation value of the single machine state is greater than the preset deviation threshold, then the beat timestamp and the running action cycle corresponding to the single machine state deviation value are extracted, and the beat timestamp and the running action cycle are combined and encapsulated to obtain the current running beat.

[0079] The adjacent nodes of the current machine are retrieved by searching the machine topology map, and the process flow sequence of the adjacent machine is obtained.

[0080] The process flow sequence is subjected to sequence embedding processing to obtain a beat feature vector. The cosine similarity between the current running beat and the beat feature vector is calculated to obtain the association matching degree.

[0081] If the correlation matching degree is greater than the preset matching threshold, then the associated running beat is extracted from the beat feature vector.

[0082] The preset deviation threshold is 10. The process for obtaining this threshold is as follows: under laboratory conditions, the same model of machine operates continuously for at least 30 days under standard operating conditions. The daily single-machine status deviation value sequence is calculated using the method described above. All deviation values ​​are sorted from smallest to largest, and the 95th percentile is taken as the preset deviation threshold. This statistical lower bound ensures that only about 5% of the samples are misjudged as deviations when the machine is operating normally (in a stable condition). This avoids system overhead caused by frequently triggering subsequent cycle extraction logic due to an excessively low threshold, while also preventing the omission of genuine operational deviations due to an excessively high threshold. In this embodiment, the single-machine status deviation value of machine number 5 is 15, which is greater than the preset deviation threshold of 10, triggering the extraction operation of the current operating cycle.

[0083] The cycle time stamp is automatically written to the operation log by the machine's built-in programmable logic controller at the start of the spindle feed action. In this embodiment, the cycle time stamp is 08:15:20. The operation cycle is the time required for the machine to complete one full spindle feed to spindle retraction action, obtained by subtracting the feed start time and retraction end time recorded by the encoder. In this embodiment, the operation cycle is 45s. The cycle time stamp and operation cycle are combined and encapsulated into a binary structure according to a preset field format to obtain the current operating cycle.

[0084] The machine topology diagram is a pre-constructed directed graph data structure. Node data comes from the equipment asset inventory in the workshop Manufacturing Execution System (MES). Each machine node includes attributes such as machine number, process category, and geographical coordinates. Directed edge data comes from the material flow path table in the process planning document and historical material flow records in the MES, with the edge direction pointing from upstream machine to downstream machine. When a machine is added or removed in the workshop, or when the material flow path changes, maintenance personnel export the latest data from the MES and reconstruct the topology diagram monthly during downtime maintenance periods. The material flow frequency and the quantity of material flowed in a single flow in the edge weights are both taken from the rolling statistics of the most recent 30 calendar days.

[0085] The edge weight is determined by the material flow frequency between the two machines. (Unit: times / day) and average quantity of materials per transaction (Unit: piece) According to the preset fusion formula The weighted average is obtained, where min-max(·) represents the normalization of this variable for all adjacent machines in the workshop. As a preset weighting coefficient, this embodiment takes... The weighting coefficient is determined based on maximizing the correlation coefficient between the fused edge weights and the actual coupling degree of material flow under no less than 30 typical working conditions, with a correlation coefficient not lower than 0.85. Based on the machine adjacency graph used as the skeleton of the perception fusion view in step S11, this step further employs a machine topology graph for rigorous adjacent node retrieval. In this embodiment, starting from the current machine node (CNC machine tool number 5) with a preset threshold of adjacent machine number N=2, all its directly downstream nodes are retrieved along the directed edges to obtain two adjacent machines: polishing machine tool number 6 and cleaning machine tool number 7. These are sorted by edge weight from largest to smallest, and the top two are retained as valid adjacent machines for subsequent processing (edge ​​weight 6 is 0.82, and edge weight 7 is 0.45, both are retained). When the number of candidate adjacent machines is less than N, all candidate machines are selected; when the number of candidate machines is greater than N, the top N are selected. If the current machine has no downstream nodes in the topology graph (e.g., a machine at the end of the production line), or the process flow sequence of all adjacent machines is empty (e.g., a newly commissioned machine), then skip the associated operating cycle extraction step and directly proceed to step S13 to calculate the difference between the current operating cycle and the standard operating cycle. The threshold N for the number of adjacent machines is determined by statistically analyzing the distribution of the number of directly downstream machines of the same model that have a real process connection relationship under typical working conditions during the commissioning and debugging phase, and taking the 90th percentile of this distribution as the N value, so as to control the computational cost of a single search while covering the vast majority of associated scenarios.

[0086] The process flow sequence is a sequence formed by arranging all discrete actions performed by adjacent machines in the current processing cycle in chronological order. It includes six discrete actions: material receiving, fixture locking, polishing start, polishing end, fixture release, and material output. Each action has a start timestamp and an end timestamp. The programmable logic controller built into the adjacent machines writes the operation log and synchronizes it to the edge server in real time.

[0087] The sequence embedding process is implemented using a simple feature engineering approach combining action dictionary lookup and timestamp normalization. The action dictionary is a key-value mapping table pre-established during the commissioning phase. The keys are discrete action names (e.g., material receiving, fixture locking, polishing start, polishing end, fixture release, material output, etc., 6 categories), and the values ​​are preset action cycle baseline values ​​(in seconds) corresponding to each action. The action dictionary is entered and stored by on-site engineers based on the standard action times given in the equipment process manual, and can be directly retrieved from the table during runtime. The maximum allowable cycle of 100 seconds is the measured maximum value of the action cycle for all discrete actions of all machines in the workshop under standard operating conditions, obtained from statistical analysis of at least 30 days of historical operating data collected during the commissioning phase, and used as the normalization baseline value for sequence embedding processing.

[0088] The core action (i.e., polishing start action) corresponding to the current operating cycle is extracted from the process flow sequence of adjacent machines. The action cycle of 52s is normalized to the 0-1 range according to the preset maximum allowable cycle of 100s, resulting in 0.52 as the first dimension component. The start timestamp of the action is normalized to the 0-1 range according to the daily working time benchmark min-max, resulting in 0.42 as the second dimension component. The daily working time benchmark min is the fixed start time of the first shift of the workshop at 08:00:00, and max is the fixed end time of the last shift at 20:00:00, which is preset and stored in the workshop schedule. The two normalized values ​​are concatenated in a predetermined order to obtain a 2D cycle feature vector. The cycle feature vector represents the cycle characteristics of adjacent machines from two dimensions: cycle length and time distribution.

[0089] Since the current operating cycle itself is a binary structure containing the cycle timestamp and the operating action cycle, it is not in the same representation space as the cycle feature vector. Therefore, the current operating cycle needs to be represented in the same space first: normalize the operating action cycle of 45s in the current operating cycle to the interval of 0 to 1 according to the same maximum allowable cycle of 100s to obtain 0.45 as the first dimension component. Normalize the cycle timestamp 08:15:20 to the interval of 0 to 1 according to the working hour benchmark of the day (min=08:00:00, max=20:00:00) min-max to obtain approximately 0.18 as the second dimension component. Thus, we obtain the 2-dimensional representation vector of the current operating cycle (0.45, 0.18). The cosine similarity calculation involves finding the cosine of the angle between two 2D vectors, summing the products of the two vectors for each component dimension as the numerator, multiplying the square root of the sum of the squares of the two vectors for each component dimension as the denominator, dividing the numerator by the denominator, multiplying the result by 100, and rounding down to the nearest integer. This yields the correlation matching degree, ranging from 0 to 100. In this embodiment, the ratio of the beat feature vector (0.52, 0.42) to the current running beat representation vector (0.45, 0.18) calculated using the formula is 0.3096 for the numerator and 0.3240 for the denominator, resulting in a ratio of approximately 0.9556. Multiplying this by 100 and rounding down yields 95, which is the correlation matching degree of 95 in this embodiment.

[0090] The preset matching threshold is 80. The acquisition process is as follows: During the workshop commissioning and debugging phase, for sample pairs in historical operating data, a standard deviation σ of the cycle time deviation between two machines within a 48-hour continuous operating window is used as a positive example label indicating the existence of genuine cycle time linkage; a standard deviation σ of σ greater than or equal to 2.0s is used as a negative example label indicating the absence of genuine cycle time linkage (samples between 1.0s and 2.0s are excluded from the calibration). Fifty pairs of positive and negative samples are collected respectively, and their correlation matching degree is calculated. Receiver operating characteristic curves are plotted for both distributions, and the Youden index (true positive rate minus false positive rate) is calculated. The correlation matching degree corresponding to the maximum value of the Youden index is taken as the preset matching threshold. The statistical result is 80, corresponding to a true positive rate of 0.86 and a false positive rate of 0.12. When the correlation matching degree is greater than this preset matching threshold, it is determined that there is a genuine cycle time linkage relationship between the current machine and its adjacent machines. The associated operating cycle time of the adjacent machines within the corresponding time window is then reverse-analyzed from the cycle time feature vector.

[0091] The specific reverse parsing method is as follows: First, query the action dictionary corresponding to the first dimension component 0.52 after inverse normalization (i.e., 0.52 × 100s = 52s). Then, find the discrete action whose action cycle is closest to 52s from the process flow sequence of the adjacent machine (i.e., polishing start, consistent with the core action recognition rule used in the embedding processing stage). Extract the running cycle of this action, 52s, as the associated running beat. When multiple action cycles have the same closeness to 52s, prioritize the action that is later in the time sequence of the process flow sequence. When the association matching degree is less than or equal to the preset matching threshold, it is determined that there is no real beat linkage between the two machines. Skip the associated running beat extraction step and directly enter step S13, but only calculate the difference between the current running beat and the standard running beat (this is a simplified branch; this embodiment will describe the normal branch that triggers extraction as an example).

[0092] In step S13, the difference between the current running cycle, the associated running cycle, and the preset standard running cycle is calculated to obtain the cycle time loss, including:

[0093] The current running beat, the associated running beat, and the preset standard running beat are aligned on the time axis to obtain a time offset sequence;

[0094] Extract process flow delays from the time offset sequence, and construct a cycle time difference matrix based on the process flow delays;

[0095] The beat difference matrix is ​​subjected to feature extraction to obtain the out-of-step feature vector. The out-of-step feature vector is then mapped to nodes in a preset multi-machine collaborative graph to obtain the beat out-of-step amount.

[0096] The preset standard operating cycle time is retrieved from the production process database. For each machine processing link, the ideal cycle time parameters are pre-calibrated during the process planning stage. In this embodiment, the standard operating cycle time of machine number 5 is 40s. The time axis alignment process is implemented using a simple method of directly subtracting timestamps. The current operating cycle time (45s), the associated operating cycle time (52s), and the standard operating cycle time (40s) are respectively expanded into discrete timestamp sequences according to the time sequence. They are paired one by one according to the preset process nodes (e.g., material receiving, fixture locking, polishing start, etc., three core processes). For each process node, the current operating cycle timetamp is subtracted from the standard operating cycle timetamp to obtain the first offset. The associated operating cycle timetamp is subtracted from the current operating cycle timetamp to obtain the second offset. The six offsets of the three process nodes are concatenated in time sequence to obtain the time offset sequence. Taking the material handover process as an example, the action timestamp of machine No. 5 is 10s, the response timestamp of machine No. 6 is 15s, and the standard timestamp is 12s. After subtraction, the first offset is -2s (that is, machine No. 5 is 2s ahead of the standard) and the second offset is +5s (that is, machine No. 6 lags behind machine No. 5 by 5s), which together constitute the time offset subsequence of this node.

[0097] The out-of-step feature vector is mapped to nodes in a pre-defined multi-machine collaborative graph to obtain the cycle time out-of-step amount. Specifically, the time offset sequence is traversed, and all process flow delay values ​​greater than 0 are extracted. A delay greater than 0 indicates that the actual triggering of the process node is later than the standard timing, which is a key factor causing cycle time out-of-step. A delay less than or equal to 0 is considered as early or on time and is uniformly recorded as 0. The extracted process flow delays are arranged in two dimensions according to the process dimension (e.g., material receiving, fixture locking, polishing start, etc., three core processes) and the machine dimension (numbered 5, numbered 6), constructing a 3-row × 2-column cycle time difference matrix. Each element in the matrix is ​​the process flow delay value (in seconds) of the corresponding process and the corresponding machine. The cycle time difference matrix intuitively reflects the time lag of upstream and downstream machines in different discrete actions.

[0098] Subsequently, Principal Component Analysis (PCA) was used to extract features from the beat difference matrix. The PCA projection matrix was pre-trained offline: during the commissioning phase, at least 1000 sets of historical beat difference matrix samples (covering typical operating conditions such as light load, heavy load, and automatic start-stop switching) were collected from the same model of machine over a period of at least 30 days. Each 3×2 beat difference matrix was flattened into a 6-dimensional original vector to form a training set. A 6×6 covariance matrix was calculated on the training set, and eigenvalue decomposition was performed. The top k eigenvectors, sorted from largest to smallest, were used to form the projection matrix. The criterion for k was that the sum of the top k eigenvalues ​​accounted for a proportion of the sum of all eigenvalues, i.e., the cumulative variance contribution rate was greater than or equal to 90%. In this embodiment, the cumulative contribution rate of the top 3 eigenvalues ​​was 92.6% (meeting the threshold), therefore k=3, and the projection matrix dimension was 6×3. During online operation, the real-time 3×2 beat difference matrix only needs to be flattened into a 6-dimensional original vector and multiplied with the pre-trained projection matrix to obtain a 3-dimensional out-of-step feature vector, where the physical unit of each component of the vector is s. The preset multi-machine collaborative graph is a pre-constructed weighted undirected graph with all processing machines in the workshop as nodes and the beat correlation degree (value from 0 to 1, dimensionless) between two machines as edge weights. The beat correlation degree is the absolute value of the correlation coefficient obtained by performing Pearson correlation analysis on historical operating beat data during the commissioning phase. In this embodiment, the beat correlation degree between machines numbered 5 and 6 is 0.78.

[0099] The current out-of-step feature vector is mapped to nodes by first-order neighbor aggregation along the edges of the collaborative graph, starting from the current machine node. To avoid the result of non-unique beat out-of-step amount due to taking different machine perspectives in multi-machine scenarios, this embodiment specifies that the current machine that triggers the current S13 step (i.e., the machine whose single-machine state deviation value exceeds the standard, which is machine number 5 in this embodiment) is the unique aggregation starting point: the out-of-step feature vector of the current machine and the out-of-step feature vectors of all its neighboring machines are weighted and summed according to the edge weights (i.e., ∑(wij×vi), where vi is the out-of-step feature vector of neighboring machine i and wij is the edge weight between the current machine and neighboring machine i). The resulting 3-dimensional vector is then subjected to the L2 norm along the time dimension (i.e., the square root of the sum of squares of each component), which yields a scalar value, defined as the beat out-of-step amount. Since the unit of each component before aggregation is s, and the L2 norm operation does not change the dimension, the physical unit of the beat loss is s. In this embodiment, the beat loss output after node mapping is 18s. The larger the value, the more serious the deviation of the beat from the standard timing of multiple machines.

[0100] In step S14, the initial feed rate and initial cooldown time are extracted based on the cycle time loss. The initial feed rate and initial cooldown time are adjusted through cross-coupling to obtain the target feed rate and target cooldown time. A synchronized cluster timing sequence is then generated based on the target feed rate and target cooldown time. This includes:

[0101] Based on the cycle step loss, extract the initial feed rate of the current machine and the initial cooling time of the adjacent machine;

[0102] The difference coupling calculation is performed between the current operating cycle time of the current machine and the associated operating cycle time of the adjacent machine to obtain the multi-machine synchronization coupling error;

[0103] A cross-coupling control algorithm is adopted to reduce the initial feed rate based on the multi-machine synchronous coupling error to obtain the target feed rate, and to extend the initial cooling time based on the reverse compensation component of the multi-machine synchronous coupling error to obtain the target cooling time.

[0104] A feedback matrix is ​​constructed based on the target feed rate and the target cooling time. The feedback matrix is ​​then encapsulated using a protocol to obtain control instructions. These control instructions are then sent to the programmable logic controller to obtain the synchronized cluster timing.

[0105] When the cycle time lag is 18 seconds (i.e., there is an 18-second cycle time lag between CNC machine tool No. 5 and polishing machine No. 6), the initial operating parameters corresponding to the current batch are collected through the equipment data interface: the initial feed rate of CNC machine tool No. 5 is 120 mm / s (read from the feed command F value in the CNC system's built-in machining program), and the initial cooling time of polishing machine No. 6 is 45 seconds (read from the M command in the polishing program). These initial parameters reflect the default operating state of the machine tools when receiving a new task assignment.

[0106] The differential coupling calculation is achieved by directly subtracting the associated operating cycle from the current operating cycle and adding a preset buffer: the difference between the associated operating cycle of 52s and the current operating cycle of 45s is 7s, which is taken as the basic coupling deviation; considering actual operating conditions such as equipment inertia and electrical response delay, a preset process switching buffer time of 8s is added to the basic deviation. The 8s buffer time is obtained by collecting no less than 30 sets of time delay samples from the issuance of the command to the actual steady state of the servo motor under load conditions for the same model of machine during the commissioning and debugging phase (covering typical operating conditions such as light load, heavy load, and automatic start-stop switching), and taking the arithmetic mean after removing abnormal samples exceeding 3σ (sample standard deviation is 1.2s, coefficient of variation is 15%, statistically robust). That is, the final multi-machine synchronization coupling error of 15s is obtained, which represents the time dimension of the polishing machine that needs to be delayed in starting.

[0107] The cross-coupling control algorithm is a multi-axis cooperative control method. Its core is to simultaneously distribute the synchronization coupling error of multiple machines to the local controllers of each machine for cooperative correction. Specifically, in this embodiment, the 15s synchronization coupling error of the multiple machines is decomposed according to a preset coupling distribution coefficient. The speed compensation component of the current machine (number 5) is Δv = Kv × E, where Kv is the speed coupling coefficient, with units of mm / (s). 2 The physical meaning of Δc is the amount of feed rate reduction corresponding to each unit of synchronous coupling error; the cooling time compensation component of the adjacent machine (number 6) is Δc=Kc×E, where Kc is the cooling coupling coefficient, the unit is dimensionless (s / s), and the physical meaning is the extension ratio of cooling time corresponding to each unit of synchronous coupling error.

[0108] The coupling coefficients Kv and Kc were obtained by conducting at least 30 sets of step response experiments on the same model of machine under laboratory conditions (synchronous coupling error with a step amplitude of 5-15s, operating under load). The optimization objective was to eliminate the synchronous coupling error in less than 2s and reduce the overshoot to less than 10%. A grid search method was used to optimize the results within the range Kv ∈ [0.5, 1.5] mm / (s). 2 The optimal combination is searched within the range of Kc∈[0.5, 1.5] with a step size of 0.1, and finally Kv=1.0mm / (s) is calibrated. 2 Kc = 1.0. Calculated using the above formula: Δv = 1.0 × 15 = 15 mm / s, Δc = 1.0 × 15 = 15 s, therefore the target feed rate is 120 mm / s − 15 mm / s = 105 mm / s; the target cooling time is 45 s + 15 s = 60 s. The allowable deviation range for the approximate 1:1 coupling relationship is ±5% (i.e., the velocity coupling coefficient is between 0.95 and 1.05 mm / (s)). 2 The deviation range is between 0.95 and 1.05 (with cooling coupling coefficients between 0.95 and 1.05), and this deviation range is determined by the 95% confidence interval of the above 30 sets of step response experiments.

[0109] A 2D feedback matrix is ​​constructed using the target feed rate of 105 mm / s and the target cooling time of 60 s as state elements. The rows of the matrix represent machine tool numbers, and the columns represent parameter types. In this embodiment, the matrix elements are the two valid elements: the feed rate of machine tool number 5 (105 mm / s) and the cooling time of machine tool number 6 (60 s). The remaining positions are filled with 0 to indicate that the corresponding machine tool does not adjust this parameter. The feedback matrix is ​​then encapsulated to obtain control commands. The encapsulation uses the binary frame format specified by the workshop fieldbus protocol (e.g., Profinet). The encapsulated content includes a frame header, machine tool number, parameter type field, parameter value field, CRC check field, and frame tail. The control commands are sent to the workshop programmable logic controller (PLC). After parsing the commands, the PLC outputs updated setpoints to the servo drivers and coolant control valves of the corresponding machine tools. Machine tool number 5 performs a feed action at 105 mm / s, and polishing machine tool number 6 performs a cooling action at 60 s, thus obtaining the synchronized cluster timing.

[0110] In step S15, if the misalignment between the cluster timing and the preset standard timing is greater than a preset misalignment threshold, a state correction command is generated to align the timing of the multiple machines, resulting in a multi-machine collaborative control outcome. This includes:

[0111] The synchronized cluster timing sequence is compared with the preset standard timing sequence to obtain the timing deviation sequence;

[0112] The time-series deviation sequence is reduced in dimensionality to obtain the core misalignment amount. If the core misalignment amount is greater than a preset misalignment threshold, the computer station state compensation matrix is ​​calculated based on the core misalignment amount.

[0113] A command sequence is constructed based on the machine state compensation matrix, and the command sequence is converted to a protocol to obtain a state correction command.

[0114] The response delay sequence of each machine is extracted according to the state correction command. An alignment compensation matrix is ​​calculated based on the maximum delay value in the response delay sequence. The alignment compensation matrix is ​​used to modulate the underlying drive signal to obtain a modulated drive signal. The operation actions of multiple machines are time-aligned using the modulated drive signal to obtain the multi-machine collaborative control result.

[0115] For the assembly line operation of CNC machine tool No. 5 and polishing machine No. 6, the cluster timing data of each machine after synchronization adjustment is acquired (written into the operation log in real time by the machine's built-in programmable logic controller when an action is triggered). Simultaneously, a preset standard timing sequence (an ideal time node sequence pre-calibrated during the process planning stage) is retrieved from the workshop manufacturing execution database. The acquired actual timing nodes and standard timing nodes are paired according to a preset action type classification system, which includes five core process actions: material receiving, fixture locking, main machining (including cutting feed or polishing start), coolant switching, and material output. The timestamps of each pair of actions are subtracted to obtain a timing deviation sequence containing multiple time differences. This timing deviation sequence reflects the overall time offset of the current cluster during the multi-process flow.

[0116] The original time-series deviation sequence is constructed in the following way: 5 types of core process actions × 2 machines × 5 time window sampling points (that is, backtracking 5 processing cycles in the current processing cycle and taking a set of time deviations for each action in each cycle), so a total of 5 × 2 × 5 = 50 dimensions are formed. The physical meaning of each dimension is the time deviation value of a certain action of a certain machine in a certain historical processing cycle relative to the standard time sequence, and the unit is seconds.

[0117] The PCA projection matrix was pre-trained offline. During the commissioning phase, at least 1000 historical time-series deviation sequence samples (covering typical operating conditions such as light load, heavy load, and automatic start-stop switching) of the same model of machine were collected over at least 90 days. Each sample was a 50-dimensional vector. A 50×50 covariance matrix was calculated for the N×50 sample matrix formed by all samples, and eigenvalue decomposition was performed. The top k eigenvectors were sorted from largest to smallest to form the projection matrix. The criterion for k was that the cumulative variance contribution rate of the top k eigenvalues ​​was greater than or equal to 85%. In this embodiment, the variance contribution rate of the first principal component was 47.3% (retaining this single principal component can explain nearly half of the total variance of the original 50-dimensional deviation sequence, which is sufficient to characterize the main time offset features), the cumulative contribution rate of the first three principal components was 82.1%, and the cumulative contribution rate of the first four principal components was 89.5%. In this embodiment, only the first principal component is retained as the core misalignment quantity (trading the lowest dimensional compact representation within the allowable variance loss range of engineering). During runtime, the inner product of the real-time 50-dimensional time-series bias sequence and the pre-trained first principal component vector yields a scalar value, defined as the core misalignment quantity, with physical units of seconds (s). In this embodiment, the output value is 8 seconds. The core misalignment quantity is a scalar value that quantifies the main time offset characteristics causing poor connection between the CNC machine tool and polishing machine processes.

[0118] The preset misalignment threshold is 5. Under laboratory conditions, time-series deviation data of the same model of machine was collected and operated continuously for at least 30 days under standard operating conditions. The core misalignment amount was calculated daily. All core misalignment amounts were sorted from smallest to largest, and the 95th percentile (statistical lower bound) was taken. Simultaneously, on-site engineers manually labeled the daily data to determine whether active correction was needed. The smaller of the minimum core misalignment amount of the manually labeled samples requiring correction and the aforementioned statistical lower bound was taken and ultimately labeled as 5. This threshold ensures that only about 5% of samples are misjudged as misaligned during normal machine operation, avoiding frequent correction triggers that could cause system jitter. In this embodiment, the core misalignment amount 8 is greater than the preset misalignment threshold 5, triggering the compensation calculation mechanism.

[0119] The calculation of the machine state compensation matrix uses a pre-established linear mapping table between misalignment and compensation value, which is then used for lookup and linear interpolation. This mapping table was pre-calibrated through a gridded disturbance experiment during the load testing phase of the same model of machine. Specifically, the calibration method involves setting 16 disturbance nodes within a 5-20s range with a step size of 1 second, using the core misalignment as the input variable. For each node, a fine-tuning coefficient disturbance of 0% to +20% is applied to the spindle speed (reference value 8000 rpm) of machine number 5, and a correction value disturbance of 0% to +40% is applied to the coolant spray pressure (reference value 2.5 MPa) of machine number 6. The evaluation index is the disturbance amount required for the core misalignment to converge to within 1 second after the disturbance is collected again. Each disturbance experiment is repeated 5 times, and the average value is taken as the optimal compensation value corresponding to that node.

[0120] Typical nodes in the mapping table are as follows: when the core misalignment is 5s, the machine state compensation matrix is ​​[1.00, 1.00] (i.e., no compensation is needed); when the core misalignment is 10s, the machine state compensation matrix is ​​[1.08, 1.17]; and when the core misalignment is 15s, the machine state compensation matrix is ​​[1.17, 1.33]. The physical basis for the different compensation growth rates in the two dimensions is that the spindle speed is more sensitive to cycle time deviation than the coolant pressure (i.e., a small adjustment in speed can significantly change the cutting cycle time), while the coolant pressure is less sensitive to cycle time deviation, thus requiring a larger correction range to produce an equivalent cycle time compensation effect. This difference in sensitivity is determined by actual measurements of the speed-cycle time gain coefficient (approximately 0.3s / rpm) and pressure-cycle time gain coefficient (approximately 0.1s / MPa) of the same model of machine equipment. When the measured core misalignment exceeds the upper limit of the calibration range by 20 seconds, compensation is performed using the boundary node of the calibration range (i.e., the compensation matrix corresponding to the core misalignment of 20 seconds), and at the same time, a system alarm is triggered to prompt maintenance personnel to intervene and investigate the abnormal working condition.

[0121] When the measured core misalignment is between two lookup table nodes (for example, core misalignment 8 in this embodiment is between 5 and 10), the corresponding compensation value is obtained by linear interpolation. The first dimension (spindle speed fine-tuning coefficient of machine tool No. 5) is calculated as 1.00 + (1.08 − 1.00) × (8 − 5) / (10 − 5) ≈ 1.05; the second dimension (coolant spray pressure correction value of machine tool No. 6) is calculated as 1.00 + (1.17 − 1.00) × (8 − 5) / (10 − 5) ≈ 1.10. Therefore, the machine tool status compensation matrix output in this embodiment is [1.05, 1.10], that is, the spindle speed of machine tool No. 5 is multiplied by a fine-tuning coefficient of 1.05, and the coolant spray pressure of machine tool No. 6 is corrected to 1.10 times the original value. The two elements in the machine status compensation matrix are arranged sequentially according to the field format of machine number, parameter type, and compensation value to construct an instruction sequence. The instruction sequence is then encapsulated into binary frames using the same fieldbus protocol as in S14 to obtain the status correction instruction. The status correction instruction is parsed by the programmable logic controller and sent to the servo driver and coolant control valve of the corresponding machine, forcing the machine to update its operating parameters according to the compensation value.

[0122] For the spindle speed fine-tuning command of machine tool No. 5 and the pressure correction command of machine tool No. 6, the action node sequence from receiving the command to the actual execution of the servo motor of each machine tool is extracted using a workshop industrial network packet capture tool. The difference between the command issuance timestamp and the actual response timestamp in the action node sequence is processed to obtain a response delay sequence containing multiple machine tool delay characteristics. In this embodiment, the response delay of machine tool No. 5 is 12ms and the response delay of machine tool No. 6 is 28ms. The maximum delay value of 28ms in the response delay sequence is extracted and compared with the pre-stored delay threshold of 15ms. The delay threshold of 15ms is obtained in a similar way to the misalignment threshold in S15, by taking the 95th percentile of the statistical data of the response characteristics of the machine tools on site.

[0123] In this embodiment, the maximum delay value of 28ms is greater than the delay threshold of 15ms, triggering the reverse derivation mechanism. The calculation of the alignment compensation matrix adopts linear reverse derivation: for each machine, the difference between the maximum delay value and the actual response delay of the machine is defined as the delay compensation amount at the time of instruction issuance for that machine. In this embodiment, the alignment compensation matrix is ​​[16ms, 0ms], that is, machine number 5 needs to delay the instruction issuance time by 16ms, while machine number 6 does not need to be adjusted (it is itself the machine corresponding to the maximum delay). The alignment compensation matrix is ​​used to modulate the underlying drive signal: the delay compensation amount is injected into the pulse width modulation (PWM) module of the corresponding machine in the programmable logic controller.

[0124] It should be noted that the PWM carrier frequency is 20kHz, the duty cycle adjustment range is 0% to 100%, and the counter bit width is 16 bits (i.e., the PWM period resolution is 65535 ticks, corresponding to a 50μs period, 1 tick ≈ 0.76ns). The conversion relationship between delay compensation amount from ms to PWM counter ticks is: tick = delay compensation amount (ms) × 10 -3 ×PWM carrier frequency (Hz) = 16 × 10 -3 ×20000 = 320 ticks, meaning the instruction triggering time of machine #5 is delayed by 320 ticks. The PWM module receives a 16-bit unsigned integer delay compensation value (range 0 to 65535) and writes it to the corresponding machine's PWM channel through the programmable logic controller's register interface. The trigger phase of the original drive signal is reconstructed by adding the delay compensation amount to the original triggering time to obtain a modulated drive signal. By sending the modulated drive signal, the cutting feed action of machine #5 and the grinding action of machine #6 are forced to start under the same absolute time reference, completing the timing alignment of the multi-machine operation actions and ultimately obtaining the multi-machine collaborative control result.

[0125] In summary, this invention discloses an intelligent control method for multi-machine collaboration, comprising: fusing environmental change data and single-machine status data of multiple machines on an industrial production line to obtain global perception information; extracting the temporal features of the global perception information to obtain a single-machine status deviation value; if the single-machine status deviation value is greater than a preset deviation threshold, extracting the current operating cycle time; retrieving adjacent machines through a machine topology map; performing sequence embedding processing on the process flow sequences of adjacent machines to obtain associated operating cycle times; calculating the difference between the current operating cycle time, the associated operating cycle time, and a preset standard operating cycle time to obtain cycle time step loss; extracting the initial feed speed and initial cooling time based on the cycle time step loss; adjusting the initial feed speed and initial cooling time through cross-coupling to obtain a target feed speed and target cooling time; generating a synchronized adjusted cluster timing sequence based on the target feed speed and target cooling time; if the misalignment between the cluster timing sequence and the preset standard timing sequence is greater than a preset misalignment threshold, generating a status correction command to perform timing alignment on the multiple machines to obtain a multi-machine collaborative control result. This invention can realize dynamic prediction and adaptive coordinated control of the operating cycle of multiple machines.

[0126] The second embodiment of the present invention provides a multi-machine collaborative intelligent control system, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the multi-machine collaborative intelligent control method as described above.

[0127] It should be noted that the multi-machine collaborative intelligent control system provided in this embodiment of the invention is used to execute all the process steps of the multi-machine collaborative intelligent control method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0128] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0129] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for intelligent control of multiple machines working together, characterized in that, include: By integrating environmental change data and individual machine status data from multiple machines on an industrial production line, global perception information is obtained, and the temporal features of the global perception information are extracted to obtain the individual machine status deviation value. If the deviation value of the single machine status is greater than the preset deviation threshold, the current running cycle is extracted, the adjacent machines are obtained by searching the machine topology map, and the process flow sequence of the adjacent machines is processed by sequence embedding to obtain the associated running cycle. The difference between the current operating cycle, the associated operating cycle, and the preset standard operating cycle is calculated to obtain the cycle step loss amount. The initial feed rate and initial cooling time are extracted based on the cycle step loss amount. The initial feed rate and the initial cooldown time are adjusted by cross-coupling to obtain the target feed rate and the target cooldown time. The synchronized cluster timing is then generated based on the target feed rate and the target cooldown time. If the misalignment between the cluster timing and the preset standard timing is greater than the preset misalignment threshold, a state correction command is generated to align the timing of the multiple machines and obtain the collaborative control result of the multiple machines.

2. The intelligent control method for multi-machine collaboration according to claim 1, characterized in that, The integrated environmental change data and individual machine status data from multiple machines on the industrial production line are used to obtain global perception information, including: The environmental change data and the single-machine state data are subjected to time offset compensation processing to obtain an aligned environment sequence and an aligned state sequence; The alignment environment sequence and the alignment state sequence are concatenated to obtain an initial fused dataset. The initial fused dataset is then filled in to obtain a complete fused dataset. The complete fusion dataset is processed to construct a view, resulting in a perceptual fusion view. Global perceptual information is then obtained from the perceptual fusion view.

3. The intelligent control method for multi-machine collaboration according to claim 1, characterized in that, The step of extracting the temporal features of the global perception information to obtain the single-machine state deviation value includes: The global perception information is sequence-segmented using a long short-term memory network to obtain initial temporal features; The initial time series features are nonlinearly mapped to obtain deep time series features; If the feature dimension of the deep temporal feature is greater than a preset dimension threshold, then principal component extraction is performed on the deep temporal feature to obtain the standard temporal feature; The state residual is calculated by comparing the standard timing characteristics with the pre-established deviation benchmark to obtain the single-machine state deviation value.

4. The intelligent control method for multi-machine collaboration according to claim 1, characterized in that, If the deviation value of the single machine status is greater than a preset deviation threshold, then the current operating cycle is extracted, and adjacent machines are retrieved through the machine topology map, including: If the deviation value of the single machine state is greater than the preset deviation threshold, then the beat timestamp and the running action cycle corresponding to the single machine state deviation value are extracted, and the beat timestamp and the running action cycle are combined and encapsulated to obtain the current running beat. The adjacent machines are obtained by searching the adjacent nodes of the current machine through the machine topology graph.

5. The intelligent control method for multi-machine collaboration according to claim 1, characterized in that, The step of embedding the process flow sequences of adjacent machines to obtain the associated operating cycle time includes: Obtain the process flow sequence of adjacent machines; The process flow sequence is subjected to sequence embedding processing to obtain a beat feature vector. The cosine similarity between the current running beat and the beat feature vector is calculated to obtain the association matching degree. If the correlation matching degree is greater than the preset matching threshold, then the associated running beat is extracted from the beat feature vector.

6. The intelligent control method for multi-machine collaboration according to claim 1, characterized in that, The step of calculating the difference between the current running cycle, the associated running cycle, and the preset standard running cycle to obtain the cycle time loss includes: The current running beat, the associated running beat, and the preset standard running beat are aligned on the time axis to obtain a time offset sequence; Extract process flow delays from the time offset sequence, and construct a cycle time difference matrix based on the process flow delays; The beat difference matrix is ​​subjected to feature extraction to obtain the out-of-step feature vector. The out-of-step feature vector is then mapped to nodes in a preset multi-machine collaborative graph to obtain the beat out-of-step amount.

7. The intelligent control method for multi-machine collaboration according to claim 1, characterized in that, The process of extracting the initial feed rate and initial cooldown time based on the cycle time loss, adjusting the initial feed rate and initial cooldown time through cross-coupling to obtain the target feed rate and target cooldown time, and generating a synchronized adjusted cluster timing sequence based on the target feed rate and target cooldown time includes: Based on the cycle step loss, extract the initial feed rate of the current machine and the initial cooling time of the adjacent machine; The difference coupling calculation is performed between the current operating cycle time of the current machine and the associated operating cycle time of the adjacent machine to obtain the multi-machine synchronization coupling error; A cross-coupling control algorithm is adopted to reduce the initial feed rate based on the multi-machine synchronous coupling error to obtain the target feed rate, and to extend the initial cooling time based on the reverse compensation component of the multi-machine synchronous coupling error to obtain the target cooling time. A feedback matrix is ​​constructed based on the target feed rate and the target cooling time. The feedback matrix is ​​then encapsulated using a protocol to obtain control instructions. These control instructions are then sent to the programmable logic controller to obtain the synchronized cluster timing.

8. The intelligent control method for multi-machine collaboration according to claim 1, characterized in that, If the misalignment between the cluster timing and the preset standard timing is greater than a preset misalignment threshold, a state correction command is generated to perform timing alignment on multiple machines, resulting in a multi-machine collaborative control outcome, including: The synchronized cluster timing sequence is compared with the preset standard timing sequence to obtain the timing deviation sequence; The time-series deviation sequence is reduced in dimensionality to obtain the core misalignment amount. If the core misalignment amount is greater than a preset misalignment threshold, the computer station state compensation matrix is ​​calculated based on the core misalignment amount. A command sequence is constructed based on the machine state compensation matrix, and the command sequence is converted to a protocol to obtain a state correction command. The response delay sequence of each machine is extracted according to the state correction command. An alignment compensation matrix is ​​calculated based on the maximum delay value in the response delay sequence. The alignment compensation matrix is ​​used to modulate the underlying drive signal to obtain a modulated drive signal. The operation actions of multiple machines are time-aligned using the modulated drive signal to obtain the multi-machine collaborative control result.

9. A multi-machine collaborative intelligent control system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent control method for multi-machine collaboration as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the intelligent control method for multi-machine collaboration as described in any one of claims 1 to 8.

Images