Plasma cutting consumable life prediction system and prediction method

Abstract

The application relates to the technical field of industrial equipment monitoring, and discloses a plasma cutting consumable life prediction system and a prediction method. The prediction system comprises sensing, collecting, processing, identifying, characteristic, integrating, predicting, outputting and updating modules. The prediction method realizes time axis alignment of electrical and fluid parameters through multiplexing of a high-frequency sampling clock and a lagging lock of a dequeue pointer. Based on a thermodynamic equilibrium state, a working cycle is divided into an arc striking section, a steady state section and an arc collecting section, and loss micro-element integration is respectively performed. Cross-coupling coefficient matrices are calculated and generated by extracting electrode and nozzle degradation characteristics. A neural network model is used to perform residual service life prediction, and a gradient descent algorithm is used to perform closed-loop iterative optimization on weight coefficients. The application realizes time sequence synchronization of multi-physical field signals, quantifies physical loss of segmented working conditions and cross-influence between consumables, and improves the accuracy of life prediction.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment monitoring technology, specifically to a life prediction system and method for plasma cutting consumable parts. Background Technology

[0002] Plasma cutting equipment is widely used in the metal processing industry, and the electrodes and nozzles inside the cutting torch are key consumable parts. Accurately predicting the remaining service life of these consumable parts is of practical significance for ensuring cutting quality and maintaining the continuity of the production line. Most existing consumable part life assessment methods rely on a single cutting count or set fixed voltage and current thresholds for expiration alarms.

[0003] In actual cutting operations, the generation and maintenance of the plasma arc involve multiple field changes in electrical properties and fluid dynamics. When existing monitoring systems simultaneously acquire high-frequency electrical signals and fluid pressure signals, time delays occur between the acquired multi-source data due to differences in the response cycles of different sensors and data transmission links. This deviation makes it difficult for the equipment to achieve synchronous alignment of different physical quantities on the same time axis, resulting in timing misalignment of the subsequently extracted operational features. At the same time, a single complete cutting cycle consists of multiple dynamic intervals, such as arc initiation, steady-state cutting, and arc termination, and the physical erosion mechanisms experienced by consumable parts are not the same in different intervals. Existing prediction methods use a globally unified fixed empirical formula for overall loss assessment, failing to perform segmented calculations for different operating conditions, thus reducing the accuracy of basic loss data.

[0004] Furthermore, electrodes and nozzles do not deteriorate in isolation during actual consumption; changes in their physical forms can have a cross-excitation effect. Existing prediction models treat different consumables as independent objects and calculate their wear separately, failing to effectively quantify the cross-coupling relationship between components. Moreover, conventional algorithms fix prediction parameters during operation and lack a closed-loop feedback mechanism that compares historical operating data with the actual scrapping state. This causes calculation errors to accumulate continuously when the system deals with complex and ever-changing cutting conditions over a long period of time, ultimately causing the output remaining service life to deviate from the actual usable state of the consumables. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a plasma cutting consumable parts life prediction system and method, which solves the problem of low accuracy in consumable parts life prediction caused by misaligned timing of multi-physics field data acquisition, failure to assess wear in segments according to working conditions, and failure to quantify the cross-deterioration effects between components in existing plasma cutting equipment.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a plasma cutting consumable parts life prediction system, comprising:

[0008] The sensing module is used to acquire the raw analog signals of the multiphysics field;

[0009] The acquisition module is used to convert the original analog signal into digital real-time operating parameters;

[0010] A processing unit, connected to the acquisition module, integrates the following:

[0011] The preprocessing module is used to filter the real-time operating parameters and perform a first-in-first-out queue delay alignment operation on the fluid dynamics signal;

[0012] The identification module is used to perform dynamic time slicing based on working condition characteristics for a single work cycle;

[0013] The feature module is used to extract physical degradation features and generate a cross-coupling coefficient matrix;

[0014] The integration module is used to perform segmented loss accumulation and penalty feature calculation, and to summarize and generate the total loss.

[0015] The prediction module is used to call the neural network model to perform remaining lifetime inference;

[0016] The output module is used to output the predicted remaining useful life data;

[0017] The update module is used to perform reverse iterative optimization on the nozzle cross-excitation weight coefficient, the electrode cross-excitation weight coefficient, and the cross-stage splash penalty weight coefficient.

[0018] Preferably, the preprocessing module constructs a cascaded digital signal processing link in the low-frequency timing identification channel to perform traversal operations on the electrical sequence, extracts the switching frequency of the inverter main circuit as the notch center frequency to construct a band-stop notch filter, and connects a low-pass digital filter in series at the back end to output a smooth electrical feature sequence; at the same time, it constructs a high-frequency feature retention channel to output a high-frequency voltage feature sequence.

[0019] Preferably, the identification module reads the real-time back pressure data of the protective gas after alignment processing and calculates the back pressure deviation of the protective gas. At the same time, it reads the first derivative of the filtered voltage, compares the absolute value of the first derivative of the filtered voltage with the preset zero bias threshold, confirms that the moment when the plasma arc and the protective gas flow field enter the thermodynamic equilibrium state is recorded as the steady-state boundary moment, and dynamically divides a single cutting cycle into an arc initiation segment, a steady-state segment, and an arc termination segment on the time axis.

[0020] Preferably, the feature module generates a high-frequency voltage envelope sequence within the steady-state interval using a digital analytical signal construction method based on Hilbert transform, calculates the mathematical variance of the high-frequency voltage envelope sequence as an electrode degradation feature, calculates the mean absolute deviation of the protective gas back pressure as a nozzle degradation feature, and calculates the unidirectional excitation factor that aggravates the electrode arc migration caused by nozzle geometric anomalies and the unidirectional excitation factor that aggravates the burn-off effect of electrode emission anomalies on the nozzle flow channel by introducing a linear weighting mechanism, thereby generating the cross-coupling coefficient matrix.

[0021] In the above system architecture, the preprocessing module outputs smooth electrical feature sequences and high-frequency voltage feature sequences, respectively; the identification module divides the time axis based on the thermodynamic state determination of the electric arc and airflow field; and the feature module calculates the cross-excitation weight coefficients and generates a cross-coupling coefficient matrix by combining electrode degradation features and nozzle degradation features. These modules work together to complete the multi-dimensional extraction and cross-correlation of consumable component state parameters.

[0022] A second aspect of the present invention provides a method for predicting the lifespan of consumable parts cut by plasma cutting, using the aforementioned plasma cutting consumable parts lifespan prediction system, the method comprising the following steps:

[0023] The acquisition module obtains digital real-time operating parameters, including transient cutting voltage, cutting current, real-time back pressure of protective gas, and ion gas flow rate.

[0024] The preprocessing module performs multi-level feature-preserving filtering on the real-time running parameters and executes a time-delay alignment operation;

[0025] The identification module combines control commands to perform dynamic timing boundary identification for a single work cycle, dynamically dividing the time axis into the arc initiation segment, steady-state segment, and arc termination segment.

[0026] The feature module extracts electrode degradation features and nozzle degradation features, and calculates and generates a cross-coupling coefficient matrix.

[0027] The integration module performs loss infinitesimal integral calculations according to the divided time intervals;

[0028] The integration module summarizes the integral values ​​of each segment to generate the total loss, and the prediction module performs inference and prediction calculations of the remaining service life through a long short-term memory neural network model.

[0029] The output module outputs a consumable parts replacement prompt when the remaining service life is lower than the safety warning threshold, and outputs a shutdown protection command when the remaining service life is lower than the forced shutdown limit threshold.

[0030] The update module uses the gradient descent algorithm to perform closed-loop iterative optimization on the weight coefficients in the cross-coupling coefficient matrix.

[0031] Preferably, the step of the acquisition module obtaining digitized real-time operating parameters includes:

[0032] The acquisition module synchronously discretizes the analog electrical signals of transient cutting voltage and cutting current, and uses the high-frequency sampling clock for acquiring electrical parameters to convert the analog signals related to fluid dynamics. It forces the fluid discrete sequence and the electrical discrete sequence to have the same time axis coordinate density, and pushes the data into the ring data buffer for continuous storage after aligning the data according to the order of generation time.

[0033] Preferably, the preprocessing module performs multi-level feature-preserving filtering on the real-time operating parameters and executes a delay alignment operation, including the following steps:

[0034] The preprocessing module calculates the absolute time lag introduced by the digital filtering link in the data channel as the group delay parameter, and calculates the offset of discrete data points in combination with the high-frequency sampling frequency; a dedicated first-in-first-out buffer queue is planned to push the discrete fluid sequence into the buffer, and the offset of the discrete data points is used to perform a hysteresis locking operation on the dequeue pointer to force the fluid signal to generate a physical residence time and complete the time axis synchronization.

[0035] Preferably, the step of the integration module performing loss infinitesimal integral calculations according to the divided time intervals includes:

[0036] The amplitude square and the absolute value of the first derivative of each data point in the discrete sequence of cutting current are extracted, multiplied by the corresponding constant coefficients, and then summed to generate a transient erosion micro-element function to calculate the transient impact loss of the arc segment; the unidirectional excitation factor is extracted and multiplied by the electrode basic micro-element loss rate and the nozzle basic micro-element loss rate to generate a steady-state integral function to calculate the coupling cumulative loss of the steady-state segment; when the variance of the high-frequency envelope of the voltage exceeds the wander limit threshold, the nonlinear penalty integral calculation logic is activated to calculate the additional splash loss.

[0037] Preferably, the step of the prediction module performing inference prediction calculations of remaining useful life through a long short-term memory neural network model includes:

[0038] The current cumulative electrode wear, current cumulative nozzle wear, high-frequency voltage envelope variance, and mean absolute deviation of protective gas back pressure from the current cutting cycle and several previous consecutive cutting cycles are combined in a fixed order to construct a two-dimensional feature matrix sequence. After linearly mapping the values ​​using the max-min normalization algorithm, the sequence is input into the long short-term memory neural network model. The output results are obtained from the linear fully connected layer, and a non-negative truncation operation is performed to output the remaining number of executable cutting cycles for the electrode and nozzle respectively. The smaller of the two values ​​is taken as the total number of remaining executable cutting cycles for the current consumable part combination.

[0039] Preferably, the update module employs a gradient descent algorithm to perform closed-loop iterative optimization of the nozzle cross-excitation weight coefficient, the electrode cross-excitation weight coefficient, and the cross-stage splash penalty weight coefficient, including the following steps:

[0040] Extract the theoretical limit loss constant from the calibrated storage and construct a global error objective function to quantify the deviation between the theoretical cumulative total loss and the actual scrap limit physical state by combining historical cyclic data; solve the first-order partial derivatives of the nozzle cross-excitation weight coefficient, electrode cross-excitation weight coefficient, and cross-stage splash penalty weight coefficient respectively to generate a three-dimensional gradient direction vector; after performing gradient absolute value comparison and upper limit limiting, iteratively generate a new generation of nozzle cross-excitation weight coefficient, electrode cross-excitation weight coefficient, and cross-stage splash penalty weight coefficient, and overwrite them into the non-volatile memory to replace the old parameter matrix.

[0041] The aforementioned method for predicting the lifespan of consumable parts cut by plasma cutting achieves synchronization of the time axis between electrical and hydrodynamic parameters through synchronous discretization and lag locking of the outgoing pointer. Combined with the identification of the time boundary of the work cycle, the time axis is divided into the arc initiation segment, the steady-state segment, and the arc termination segment, and the corresponding infinitesimal integral calculation is performed for each segment to generate the total loss. Subsequently, the remaining number of executable cutting cycles is output by the prediction model, and the weight coefficients in the cross-coupling coefficient matrix are iteratively updated by calculating the global error objective function and partial derivatives to complete the lifespan prediction calculation.

[0042] This invention provides a system and method for predicting the lifespan of consumable parts cut by plasma cutting. It has the following beneficial effects:

[0043] 1. This invention converts analog signals related to fluid dynamics by multiplexing the high-frequency sampling clock of electrical parameters, and performs a hysteresis locking operation on the dequeue pointer by combining the calculated discrete data point offset. This data processing method forces the fluid signal to generate physical dwell time, eliminates the data acquisition time deviation between electrical signals and fluid dynamic signals, and realizes the coordinate synchronization of different physical field parameters on the same time axis.

[0044] 2. This invention performs dynamic time slicing on a single cutting cycle based on the thermodynamic equilibrium state of the plasma arc and the airflow field. The time axis is divided into arc initiation segment, steady state segment and arc termination segment, and loss infinitesimal integration is performed on each segment. The degradation characteristics of the electrode and nozzle are extracted to calculate the cross-coupling coefficient matrix. This segmented integration and cross-coupling coefficient processing mechanism quantifies the specific physical losses in different working conditions and the mutual influence between multiple consumable components, thereby improving the accuracy of total loss data calculation.

[0045] 3. This invention constructs an iterative optimization mechanism for the parameters of the prediction algorithm. By combining historical cyclic data with the calibrated limit loss constant, a global error objective function is constructed, and the first-order partial derivatives of the weight coefficients are solved to generate gradient direction vectors. Then, new weight coefficient values ​​are iteratively generated and overwritten to memory. This setting enables the cross-coupling coefficient of the system to be closed-loop corrected based on the actual cumulative loss deviation, reducing error accumulation during long-term operation and maintaining the effectiveness of lifetime prediction results. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the plasma cutting consumable parts life prediction system architecture of the present invention;

[0047] Figure 2 This is a schematic diagram of the overall process of the plasma cutting consumables life prediction method of the present invention;

[0048] Figure 3 This is a schematic diagram of the dynamic temporal boundary recognition process of the present invention;

[0049] Figure 4 This is a schematic diagram of the process for extracting degradation features and constructing the cross-coupling coefficient matrix in this invention;

[0050] Figure 5 This is a schematic diagram of the segmented loss calculation for a single cutting cycle according to the present invention, wherein a is a schematic diagram of the division of the cutting current time sequence interval, and b is a schematic diagram of the corresponding component cumulative loss change.

[0051] Figure 6 This is a schematic diagram of the remaining useful life inference and model update process of the present invention;

[0052] Figure 7 This is a schematic diagram of the time-series slicing of the multi-physics parameters of the present invention, wherein a is a schematic diagram of the cutting current changing with time, b is a schematic diagram of the voltage envelope changing with time, and c is a schematic diagram of the real-time back voltage changing with time.

[0053] Figure 8 This is a scatter plot of the distribution of degradation characteristics of multi-component cross-coupling in this invention;

[0054] Figure 9 This is a schematic diagram comparing the remaining lifetime prediction effects of different prediction strategies of the present invention. Detailed Implementation

[0055] The technical solutions in 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.

[0056] See attached document Figure 1 This invention provides a plasma cutting consumable parts life prediction system, which is set on the periphery and inside of the plasma cutting machine; the physical hardware platform and logic control architecture provided by the prediction system are used to execute the plasma cutting consumable parts life prediction method proposed in this invention.

[0057] The prediction system includes a sensing module, a data acquisition module, and a processing unit.

[0058] The sensing module is installed in the electrical circuit and auxiliary gas circuit of the plasma cutter; the sensing module includes a Hall voltage sensor, a Hall current sensor, a high-frequency pressure transmitter, and a flow sensor. The sensing module is used to acquire raw analog signals of multiphysics in real time.

[0059] The acquisition module is electrically connected to the sensing module; the acquisition module performs high-frequency sampling and analog-to-digital conversion on the analog signal output by the sensing module to generate digital real-time operating parameters.

[0060] The processing unit is connected to the acquisition module via an industrial communication bus; the processing unit integrates multiple logic execution units, including a preprocessing module, an identification module, a feature module, an integration module, a prediction module, an output module, and an update module.

[0061] The preprocessing module establishes a low-frequency timing identification channel and a high-frequency feature retention channel for electrical signals. The low-frequency timing identification channel performs multi-level digital filtering on the electrical signals and outputs a smooth electrical feature sequence. The high-frequency feature retention channel retains the high-frequency voltage components related to arc migration and performs first-in-first-out queue delay alignment on the fluid signals. The identification module performs dynamic time slicing based on operating condition characteristics for a single working cycle. The feature module extracts physical degradation features and generates a cross-coupling coefficient matrix. The integration module performs segmented loss accumulation and penalty feature calculation. The prediction module calls a neural network model to perform inference calculations for the remaining service life. The output module outputs the predicted remaining service life data and protection commands. The update module performs global reverse iterative optimization of the prediction algorithm parameters.

[0062] The aforementioned sensors, communication hardware, and logic execution unit constitute the physical execution carrier of the prediction method of this invention.

[0063] See attached document Figure 2 This invention provides a method for predicting the lifespan of consumable parts cut by plasma cutting, comprising the following steps:

[0064] S1, the acquisition module acquires digital real-time operating parameters at a preset sampling frequency. The real-time operating parameters include transient cutting voltage, cutting current, real-time back pressure of protective gas, and ion gas flow rate.

[0065] S2, the preprocessing module receives real-time operating parameters and performs multi-level feature-preserving filtering on the electrical signal, and performs delay alignment operation on the fluid dynamics signal according to the group delay parameters set by the filter to eliminate time axis calculation deviations across physical field data;

[0066] S3, the identification module reads the alignment-processed protective gas back pressure deviation and the filtered first derivative of the voltage, and combines the control command to perform dynamic timing boundary identification on a single working cycle of the plasma cutter, dynamically dividing the single cutting cycle into the arc initiation segment, steady state segment and arc termination segment on the time axis;

[0067] S4, the feature module extracts the high-frequency envelope variance of the voltage as the electrode degradation feature in the steady-state range, extracts the mean absolute deviation of the protective gas back pressure as the nozzle degradation feature, and calculates and generates the cross-coupling coefficient matrix based on the extracted features.

[0068] S5, the integration module performs loss infinitesimal integral calculations according to the divided time intervals. After calculating the transient impact loss of the arc initiation segment and the coupling cumulative loss of the steady-state segment, the high-frequency envelope variance of the voltage is used as a pre-trigger condition to perform splash additional loss calculations in the arc termination segment interval.

[0069] S6, the integration module summarizes the integral values ​​of the arc initiation segment, steady state segment and arc termination segment to generate the total loss of a single cutting cycle. The prediction module, based on the total loss and multi-dimensional operating characteristics, performs inference prediction calculations on the remaining service life of the electrode and the matching nozzle through a long short-term memory neural network model.

[0070] S7, the output module monitors the numerical status of the remaining service life data. When the calculated remaining service life is lower than the preset threshold, it outputs a shutdown protection command and a multi-component collaborative replacement prompt to the main controller of the plasma cutter.

[0071] S8, after the plasma cutter completes the physical replacement of consumable parts, the update module retrieves the complete wear sequence and actual working cycle number of the previous life cycle. With the minimization of global prediction error as the objective function, the gradient descent algorithm is used to perform closed-loop iterative optimization of the weight coefficients in the cross-coupling coefficient matrix and the cross-stage spatter penalty weight coefficients. The optimized parameters are then overwritten into the internal storage unit as the calculation benchmark for the next life cycle.

[0072] In this embodiment, step S1 provided by the present invention may include the following steps in specific implementation:

[0073] S101, the acquisition module executes the high-frequency synchronous acquisition and caching logic of multi-dimensional operating parameters. It receives the analog electrical signals extracted by the sensing module from the electrical circuit and gas pipeline of the plasma cutter, and converts the acquired analog electrical signals into discrete digital time series. In this process, the acquisition module performs synchronous discretization processing on the analog electrical signals of transient cutting voltage and cutting current through the internal analog-to-digital conversion unit.

[0074] Considering that the plasma cutting power supply undergoes high-frequency chopping through the inverter and the overall working environment is accompanied by high-frequency pulses, the analog-to-digital conversion unit uses a sampling frequency determined based on the switching frequency of the plasma cutting power supply inverter and the highest effective frequency of the voltage envelope to be retained for signal acquisition. This sampling frequency is not less than 5 times the highest effective frequency to be retained. When the inverter switching frequency is higher than the preset sampling capability of the acquisition module, the acquisition module retains the target envelope frequency band through front-end analog band limiting processing, or increases the sampling frequency accordingly to avoid signal aliasing distortion caused by undersampling and fully extract the high-frequency voltage envelope features to characterize the subsequent arc drift state.

[0075] S102, the acquisition module acquires the fluid dynamics parameters of the plasma cutter within the synchronous cycle of acquiring electrical parameters. It receives the analog voltage output from the pressure transmitter installed at the back end of the gas proportional valve and the analog current output from the flow sensor in the main gas pipeline. Based on the low-frequency slow change characteristic of the physical change rate of fluid dynamics parameters, the acquisition module reuses the high-frequency sampling clock of the aforementioned electrical parameter acquisition to convert the analog signals related to fluid dynamics and maintain the numerical filling at continuous time.

[0076] The downsampling mechanism is eliminated to force the discrete fluid sequence and the discrete electrical sequence to have completely consistent data length and time axis coordinate density, thereby directly generating discrete real-time back pressure data of protective gas and flow rate data of ion gas. As for the specific hardware circuit architecture design of the analog-to-digital conversion unit and the configuration of the multi-channel sampling clock, in this embodiment, conventional design can be carried out according to the actual selected control chip resources, and its underlying conversion circuit belongs to the well-known technology in the field.

[0077] S103, the acquisition module opens a circular data buffer in its internal storage space to continuously store the generated digital real-time running parameter sequence. This circular data buffer is used to save the data within the current sliding time window and is not used as the sole storage space for a complete single cutting cycle. For cutting cycles whose duration exceeds the sliding time window, the processing unit uses a streaming statistical accumulator and a piecewise integral accumulator to update the feature quantity and loss quantity in real time.

[0078] To address the time-span fluctuations caused by differences in the cutting trajectory during a single plasma cutting cycle, the acquisition module strictly aligns the acquired transient cutting voltage, cutting current, real-time back pressure of the protective gas, and ion gas flow rate according to the chronological order of data generation before pushing them into a circular data buffer. This circular data buffer adopts a data structure design with the beginning and end connected. When a newly entered discrete data point fills the end of the preset physical address space, the memory access pointer automatically jumps to the starting address to overwrite it. The depth and capacity of the circular data buffer are jointly calculated and determined by the memory space allocated by the system, the set length of the operation time window, and the pre-set sampling frequency.

[0079] The queue length limit of the circular data buffer is configured based on specific system operational requirements, and its configuration model satisfies the following relation:

[0080] ;

[0081] In the formula, This represents the maximum total number of discrete points of a single physical operating parameter that can be accommodated within the circular data buffer; its upper limit is limited by the available static memory capacity within the processing unit. This indicates a rounding up operation to ensure complete truncation of time-series data; The physical length of the sliding time window required for the system's operational logic; The value range is set between 0.5 seconds and 2.0 seconds to cover the typical microscopic dynamic characteristic cycle of plasma cutting; This indicates the system sampling frequency at which the acquisition module performs analog-to-digital conversion. The value range is set between 10,000 Hz and 50,000 Hz under conventional inverter power supply conditions to match the switching characteristics of conventional inverter power supplies. When a higher sampling rate is required to retain the highest effective frequency, According to the aforementioned sampling frequency determination rule, when the newly acquired discrete data points transmitted by the underlying hardware enter the circular data buffer, the acquisition module automatically covers the data node with the oldest timestamp by relying on pointer displacement, so as to keep the data volume in the buffer area always the dynamic parameter set within the latest time window under the current cutting condition.

[0082] In this embodiment, step S2 provided by the present invention may include the following steps in specific implementation:

[0083] S201, the preprocessing module extracts the discrete sequence of transient cutting voltage and cutting current from the aforementioned ring data buffer and executes multi-level digital filtering logic. The preprocessing module constructs a cascaded digital signal processing link in the low-frequency timing identification channel to perform traversal calculations on the electrical sequence. The preprocessing module extracts the switching frequency of the inverter main circuit of the underlying device as the notch center frequency to construct a band-stop notch filter. The band-stop notch filter filters out high-frequency inverter pulses and their higher harmonic components.

[0084] The preprocessing module connects a low-pass digital filter in series at the end of the band-stop notch filter link. The band-stop notch filter and the low-pass digital filter are preferably linear phase finite-length unit impulse response filters, or the total group delay of the entire filtering link is obtained by the preprocessing module through calibration. The preprocessing module sets the cutoff frequency of the low-pass digital filter in the range of 200 Hz to 500 Hz based on the spectral range of the macroscopic shift of the plasma arc. The preprocessing module outputs a smooth electrical feature sequence that retains only the envelope of the macroscopic physical state. This smooth electrical feature sequence is used for subsequent dynamic timing boundary identification. At the same time, the high-frequency feature retention channel outputs a high-frequency voltage feature sequence that is on the same time base as the smooth electrical feature sequence, which is used for subsequent calculation of the high-frequency envelope variance of the voltage.

[0085] S202, after completing the multi-level feature retention filtering process, the preprocessing module synchronously calculates the absolute time lag introduced by the digital filtering link in the data channel. This absolute time lag is the total group delay of the low-frequency time sequence recognition channel participating in the time sequence recognition. When the delay of the high-frequency feature retention channel is inconsistent with that of the low-frequency time sequence recognition channel, the preprocessing module performs timestamp compensation on the high-frequency feature retention channel so that the data of each channel enters the subsequent modules under the same time reference.

[0086] The low-pass digital filter configured in the preprocessing module is a finite-length unit impulse response filter with linear phase characteristics. When performing frequency-domain attenuation filtering on electrical signals, this filter generates a constant and quantifiable delay on the original input sequence on the time axis. The preprocessing module derives the exact group delay parameter based on the filter order configured in the underlying computing core and the actual digital sampling frequency currently being executed by the system. The preprocessing module uses the derived group delay parameter as the time axis reference correction basis for cross-physics data fusion calculation.

[0087] The specific mathematical model for calculating the group delay parameter satisfies the following relationship:

[0088] ;

[0089] In the formula, This represents the filter group delay time constant derived by the preprocessing module; This indicates the order of the finite-length unit impulse response filter configured in the preprocessing module; Set to an even integer greater than 64; This indicates the high-frequency sampling frequency at which the system performs analog-to-digital conversion.

[0090] After obtaining the aforementioned filter group delay time constant, the preprocessing module calculates the discrete data point offset by combining it with the high-frequency sampling frequency of the analog-to-digital converter. The conversion model satisfies the following relationship:

[0091] ;

[0092] In the formula, This represents the offset of the discrete data point that needs to be applied to the queue pointer; This represents the filter group delay time constant derived by the preprocessing module; This indicates the high-frequency sampling frequency at which the system performs analog-to-digital conversion.

[0093] S203, the preprocessing module performs a first-in-first-out (FIFO) queue delay alignment operation on the fluid dynamics signal according to the determined discrete data point offset. The preprocessing module plans an independent storage block in the static random access memory as a dedicated FIFO buffer queue for fluid signals. The preprocessing module continuously pushes the synchronously acquired real-time back pressure discrete sequence of protective gas and discrete sequence of ion gas flow rate into the FIFO buffer queue. The preprocessing module uses the calculated discrete data point offset to perform a corresponding hysteresis locking operation on the dequeue pointer.

[0094] The forced fluid dynamics signal generates a physical dwell time equal to the group delay time constant within the queue. The output fluid dynamics signal is shifted backward on the time coordinate system and synchronized with the filtered electrical signal on the time axis. The specific window function design method for the filter coefficients and the memory pointer transfer method of the first-in-first-out queue in the underlying operating system can be conventionally configured according to digital signal processing theory in this embodiment, which are well-known technologies in the field.

[0095] See attached document Figure 3 In specific implementations, step S3 provided by the present invention may include the following steps:

[0096] S301, the identification module reads the real-time back pressure data of the protective gas after alignment processing output by the preprocessing module, the identification module calls the current working condition standard target back pressure value set in the equipment control system, the identification module performs the difference operation between the real-time back pressure data of the protective gas and the standard target back pressure value and takes the absolute value to generate a dynamic protective gas back pressure deviation.

[0097] The identification module substitutes the mean or median of the back pressure deviation of the protective gas within a preset short window into the internally configured delay mapping function to calculate the delay compensation time for the establishment of the flow field, so as to avoid misjudgment of the steady-state boundary caused by fluctuations of a single sampling point. This delay compensation time characterizes the lag time span experienced by the gas flow field to reach thermodynamic equilibrium after the physical degradation of the geometry of the nozzle gas path of the plasma cutter.

[0098] The specific mathematical calculation model for the back pressure deviation of the protective gas satisfies the following relationship:

[0099] ;

[0100] In the formula, This indicates the deviation of the protective gas back pressure under continuous time variables calculated by the identification module. This indicates the real-time back pressure of the protective gas after time axis alignment. This indicates the standard target back pressure currently configured in the system.

[0101] The delay mapping function calculation model for delay compensation time satisfies the following relationship:

[0102] ;

[0103] In the formula, This represents the flow field delay compensation time calculated and output by the identification module; This indicates the deviation of the protective gas back pressure under continuous time variables calculated by the identification module; The value of the mapping slope constant is determined by physical calibration experiments based on the volumetric characteristics of the gas pipeline network of the specific plasma cutting system. In this embodiment, the value of the mapping slope constant ranges from 0.05 to 0.25 s / kPa.

[0104] S302, the identification module extracts the start cutting command issued by the underlying control system and records the time of its receipt as the initial zero point. The identification module simultaneously performs a first-order derivative operation on the cutting voltage data in the smooth electrical characteristic sequence output by the preprocessing module to generate the first-order derivative of the filter voltage. The identification module compares the absolute value of the first-order derivative of the filter voltage with the internally set zero bias threshold. The zero bias threshold is set according to the inherent ripple parameter of the steady-state output voltage of the cutting power supply and its value range is configured in the range of 2 volts to 10 volts per second.

[0105] When the identification module determines that the state of the absolute value of the first derivative being continuously less than the zero bias threshold has been maintained for 50 sampling periods and the current time node exceeds the sum of the initial zero point and the delay compensation time, the identification module confirms that the plasma arc and the protective gas flow field have entered a state of thermodynamic equilibrium. The identification module records the physical moment that meets the aforementioned determination conditions as the steady-state boundary moment and defines the time interval from the initial zero point to the steady-state boundary moment as the dynamic arc initiation segment.

[0106] S303, the identification module continuously monitors the operating status of the underlying control system after the end of the dynamic arc initiation segment and records the receiving time of the stop cutting command as the arc termination start time. The identification module continuously monitors the cutting current data in the smooth electrical characteristic sequence for zero crossing from the arc termination start time. When the identification module determines that the current cutting current value decays to less than or equal to the preset zero current threshold and continuously maintains the preset sampling period, it records the physical moment when the arc energy is completely reduced to zero as the arc termination end time. The identification module divides the time interval from the steady-state boundary time to the arc termination start time into a dynamic steady-state segment.

[0107] The identification module divides the time interval from the start time of arc termination to the end time of arc termination into dynamic arc termination segments. The identification module completes the time-series slicing of a single working cycle through the aforementioned physical quantity dynamic feature judgment mechanism and outputs the exact time coordinate boundaries of each stage to the subsequent modules. When no steady-state boundary is identified before the stop cutting command arrives, the identification module marks the working cycle as an abnormal short cycle or an arc ignition failure cycle and outputs an abnormal cycle flag to the integration module, so that the integration module records or removes it according to the abnormal cycle loss rule, avoiding the failure of subsequent feature extraction due to the number of sampling points in the steady-state segment being zero.

[0108] See attached document Figure 4 In specific implementations, step S4 provided by the present invention may include the following steps:

[0109] S401, the feature module receives the dynamic steady-state time coordinate boundary output by the identification module and extracts the transient cut voltage sequence obtained by the acquisition module within the exact time interval. The feature module uses a digital analytical signal construction method based on Hilbert transform to perform analytical operations on the high-frequency feature retention channel voltage signal within the extracted time period. Before the analytical operation, the voltage signal is subjected to DC removal processing and target frequency band limiting processing. Then, its magnitude value is taken to generate a high-frequency voltage envelope sequence. The feature module further calculates the mathematical variance of the high-frequency voltage envelope sequence and outputs the calculated high-frequency voltage envelope variance as the electrode degradation feature.

[0110] For steady-state segments whose duration exceeds the sliding time window of the circular data buffer, the feature module uses a recursive mean and recursive variance algorithm to update the high-frequency envelope variance of the voltage online, without having to cache the complete steady-state voltage sequence at once. This feature extraction logic is based on the physical correlation that physical erosion and depression on the surface of the electrode emitter during the operation of the plasma cutter will inevitably lead to increased arc migration. The feature module objectively quantifies the degree of physical degradation of the geometry of the electrode emitter end through the amplification process of the variance value.

[0111] The specific calculation model for the variance of the high-frequency voltage envelope satisfies the following relationship:

[0112] ;

[0113] In the formula, The variance of the high-frequency envelope of the voltage; This represents the total number of sampling points extracted within the steady-state range. The index number of the discrete sampling point; For the first The voltage high-frequency envelope amplitude corresponding to each sampling point; It is the arithmetic mean of the high-frequency envelope amplitude of the voltage within the steady-state range.

[0114] S402, the feature module synchronously extracts the protective gas back pressure deviation sequence calculated and output by the identification module in the previous step within the same dynamic steady-state time interval. The feature module performs an arithmetic mean operation on all discrete data points within the time slice to generate the average absolute deviation of the protective gas back pressure. The feature module outputs the average absolute deviation of the protective gas back pressure as a nozzle degradation feature to the subsequent modules. This extraction method combines the abnormal physical state of fluid channel resistance caused by the adhesion of liquid metal splashes or the continuous scouring and expansion of the nozzle inner hole by the high-temperature airflow. The feature module uses the degree of deviation of the absolute value of the mean from the reference zero position to quantify the physical degradation of the nozzle gas path micro-geometry.

[0115] The calculation model for the mean absolute deviation of the protective gas back pressure satisfies the following relationship:

[0116] ;

[0117] In the formula, This represents the average absolute deviation of the protective gas back pressure within the steady-state range. This represents the total number of sampling points extracted within the steady-state range. The index number of the discrete sampling point; For the first The protective gas back pressure deviation value corresponding to each sampling point.

[0118] S403, the feature module constructs a cross-coupling coefficient matrix based on the electrode degradation features and nozzle degradation features extracted in the previous step. During the operation of the plasma cutting consumable system, there is a physical mechanism of multi-component interactive damage. The feature module introduces a linear weighting mechanism to calculate the unidirectional excitation factor that aggravates the electrode arc migration due to nozzle geometric anomalies and the unidirectional excitation factor that burns the nozzle flow channel due to electrode emission anomalies. The feature module integrates the two unidirectional excitation factors into a second-order diagonal matrix as the feature amplification factor input source for subsequent infinitesimal integral operations. This second-order diagonal matrix quantifies the cross-catalytic acceleration effect of the physical morphological degradation of a single consumable on the wear rate of adjacent matching consumables.

[0119] The specific model for constructing the cross-coupling coefficient matrix satisfies the following relationship:

[0120] ;

[0121] In the formula, This is the cross-coupling coefficient matrix for multiple components; The excitation weighting coefficient is used to map the nozzle back pressure degradation characteristics to the electrode. This represents the average absolute deviation of the protective gas back pressure within the steady-state range. The excitation weighting coefficient is used to map the electrode envelope variance degradation characteristics to the nozzle. This represents the variance of the high-frequency envelope of the voltage. and These represent the backpressure degradation feature mapping weight coefficient and the envelope variance mapping weight coefficient, respectively. Both are preset to empirical constants between 0.01 and 0.1 during the system initialization phase and are subsequently optimized by the update module in combination with real cut lifecycle sequence data through closed-loop iteration.

[0122] See attached document Figure 5 In specific implementations, step S5 provided by the present invention may include the following steps:

[0123] S501, the integration module receives the dynamic arc segment time coordinate boundary output by the identification module in the previous step and extracts the corresponding discrete sequence of cutting current within the interval to perform transient impact loss integral calculation. The integration module extracts the square of the amplitude of each data point in the discrete sequence of cutting current as the direct calculation variable characterizing physical erosion, and simultaneously extracts the absolute value of the first derivative of the smoothed cutting current as the calculation variable characterizing mechanical impact.

[0124] The derivative peaks exceeding the preset physical upper limit are subjected to amplitude limiting. The integration module multiplies the two extracted calculation variables by the preset corresponding constant coefficients and then sums and superimposes them to form a transient erosion infinitesimal function. The integration module performs numerical definite integral operation on the transient erosion infinitesimal function with the system sampling period as the time step within the set time interval of the entire dynamic arc-starting segment to generate the transient impact loss of the arc-starting segment.

[0125] The calculation model for transient impact loss in the arc segment satisfies the following relationship:

[0126] ;

[0127] In the formula, The transient impact loss of the arc segment is calculated; This refers to the initial zero point of the dynamic arc segment; This is the initial moment of the steady-state boundary. The thermal erosion coefficient is defined as 1.5 × 10⁻⁶ in this embodiment. -6 Up to 3.0×10 -6 ; For a moment The corresponding continuous plasma cutting current value; The electromagnetic impact constant is defined as 0.01 to 0.05 in this embodiment. For a moment The first derivative of the change in cutting current; The independent variable is a continuous-time variable; It is a time integral infinitesimal element.

[0128] S502, the integration module receives the time coordinate boundary of the dynamic steady state segment and performs multi-component coupling cumulative loss calculation in the exact continuous interval by combining the cross-coupling coefficient matrix generated by the previous feature extraction step. The integration module extracts the product of the cutting voltage sequence and the cutting current sequence at the corresponding time to characterize the arc power and uses it as the basic micro-element loss rate of the electrode. At the same time, the integration module extracts the ion gas flow rate sequence as the basic micro-element loss rate of the nozzle.

[0129] The integration module extracts the unidirectional excitation factor of the nozzle to the electrode from the second-order cross-coupling coefficient matrix generated by the feature module, multiplies it by the aforementioned electrode basic micro-element loss rate to construct the electrode steady-state integral function, and extracts the unidirectional excitation factor of the electrode to the nozzle, multiplies it by the aforementioned nozzle basic micro-element loss rate to construct the nozzle steady-state integral function. During the time interval of the dynamic steady-state segment, the integration module performs definite integral operations on the above two integral functions respectively and continuously outputs the steady-state segment coupling cumulative loss.

[0130] The calculation model for the cumulative coupling loss in the steady-state section satisfies the following relationship:

[0131] ;

[0132] ;

[0133] In the formula, To calculate the steady-state cumulative loss of the output electrodes; This is the initial moment of the steady-state boundary. This marks the start of the arc-closing phase. The nozzle cross-excitation weighting coefficient; To protect the mean absolute deviation of the back pressure; The electrode power loss conversion constant is defined as 2.0 × 10⁻⁶. In this embodiment, the value range of the electrode power loss conversion constant is 2.0 × 10⁻⁶. -7 Up to 5.0×10 -7 ; For a moment The corresponding continuous plasma cutting current value; For a moment The corresponding continuous cutting voltage value; The independent variable is a continuous-time variable; For time integral infinitesimal elements, To calculate the steady-state cumulative loss of the nozzle output; These are the electrode cross-excitation weighting coefficients; The variance of the high-frequency envelope of the voltage; The nozzle fluid scour loss conversion constant is used. In this embodiment, the value range of the nozzle fluid scour loss conversion constant is 1.0 × 10⁻⁶. -5 Up to 4.0×10 -5 ; This represents the continuous ion gas flow rate value corresponding to the given time. The independent variable is a continuous-time variable; For the time differential element, the specific solution method for the numerical discrete integral at the microcontroller's underlying layer can be achieved in this embodiment using the conventional Runge-Kutta algorithm, which is a well-known technology in this field.

[0134] S503, the integration module reads the time coordinate boundary of the dynamic arc segment and calls the high-frequency envelope variance of the voltage generated by the feature module in the pre-calculation stage as the pre-trigger condition to perform the additional loss calculation of the splash. The integration module compares the obtained variance value with the internally preset wander limit threshold and activates the nonlinear penalty integration calculation logic when the variance value exceeds the wander limit threshold. The integration module establishes an exponential amplification penalty model with the natural constant as the base and the result of multiplying the variance value by the penalty weight as the exponent.

[0135] Before the exponential operation, threshold normalization and upper limit clipping are performed on the variance of the high-frequency envelope of the voltage to prevent the exponential term from overflowing under extreme conditions in the underlying controller. The integral module uses the exponential amplification penalty model as the integrand and performs definite integral operation to generate splash additional loss over the entire time interval of the dynamic arc-end segment. When the integral module determines that the variance value does not exceed the configured wander limit threshold, it automatically sets the splash additional loss variable to zero to avoid the numerical calculation dead zone of the underlying control system.

[0136] The specific segmented calculation model for splash additional loss satisfies the following relationship:

[0137] ;

[0138] In the formula, To calculate the additional splash loss in the output; The mathematical symbol for definite integrals; This marks the start of the arc-closing phase. This is the moment when the arc-closing phase ends; The basic penalty constant, in this embodiment, ranges from 1.5 to 3.5. The base weight of the natural logarithm; The cross-stage splash penalty weighting coefficient is, in this embodiment, the value range of the cross-stage splash penalty weighting coefficient is 0.1 to 0.5; The variance of the high-frequency envelope of the voltage in the pre-stage; The independent variable is a continuous-time variable; For time integral infinitesimal elements; The wander limit threshold set for the system is determined by high-speed camera physical observation during the system debugging phase based on the specific plasma electrode design specifications and is directly stored in the non-volatile memory unit of the underlying controller as a static comparison benchmark. In this embodiment, the wander limit threshold ranges from 15 to 30 volts squared.

[0139] See attached document Figure 6 In specific implementations, step S6 provided by the present invention may include the following steps:

[0140] S601, the integration module receives the transient impact loss of the arc-starting segment, the steady-state cumulative loss of the electrode, the additional splash loss, and the steady-state cumulative loss of the nozzle from the previous calculation output. The integration module performs arithmetic summation on the transient impact loss of the arc-starting segment, the steady-state cumulative loss of the electrode, and the additional splash loss to generate the electrode single-cycle loss characterizing a single working cycle. The integration module extracts the electrode single-cycle loss and adds it to the electrode historical total loss variable recorded in the non-volatile storage unit to update the current cumulative loss of the electrode.

[0141] The integration module synchronously uses the nozzle steady-state cumulative loss as the nozzle single-cycle loss characterizing a single working cycle, and adds it to the recorded nozzle historical total loss variable to update the nozzle current cumulative loss. At the same time, the integration module sequentially appends the various single-cycle cumulative loss values ​​generated in this cycle and the corresponding characteristic state variables to the service life data log opened in the system non-volatile memory.

[0142] Simultaneously, the basic integral term, trigger state quantity, steady-state duration, arc-ending duration, high-frequency envelope variance of voltage, mean absolute deviation of protective gas back pressure, and electrode basic loss term and nozzle basic loss term before weighting are written for subsequent parameter recalculation. The integral module uses the updated current cumulative electrode loss and current cumulative nozzle loss as the benchmark for the physical wear state of consumable parts in the current system and continuously outputs the corresponding values ​​to the subsequent prediction link.

[0143] The calculation model for the current cumulative loss of the electrode and nozzle satisfies the following relationship:

[0144] ;

[0145] ;

[0146] In the formula, For the first The current cumulative electrode wear is updated at the end of the next cutting cycle; This is the index of the discrete number of iterations in a single cutting cycle; For the first Total electrode history loss stored at the end of the next cutting cycle; This refers to the transient impact loss of the arc segment under the current cycle; This represents the steady-state cumulative loss of the electrodes under the current cycle. Adds additional losses to the splash during the current cycle; For the first The current cumulative nozzle wear is updated at the end of the next cutting cycle; For the first Total nozzle history loss stored at the end of each cutting cycle; This represents the nozzle's steady-state cumulative loss during the current cycle.

[0147] S602, the prediction module calls the pre-configured long short-term memory neural network model to perform inference calculations on the remaining service life of consumables. The prediction module combines the current cumulative loss of the electrode and the current cumulative loss of the nozzle, as well as the variance of the high-frequency envelope of the voltage and the mean absolute deviation of the back pressure of the protective gas, in a fixed order to construct a two-dimensional feature matrix sequence containing time-series evolution characteristics. The prediction module uses the maximum-minimum normalization algorithm to linearly map the values ​​of each dimension of the two-dimensional feature matrix sequence to the range of 0 to 1.

[0148] The prediction module inputs the preprocessed feature matrix sequence into the input layer of the long short-term memory neural network model and makes it flow sequentially through the hidden layer with 128 neurons and the linear fully connected layer with two output nodes. The prediction module obtains the output result from the fully connected layer and performs a non-negative truncation operation by comparing it with 0 and taking the maximum value to form the final predicted value. The two output nodes output the remaining number of executable cutting cycles for the electrode and the remaining number of executable cutting cycles for the nozzle, respectively. The prediction module further takes the smaller value of the two as the total number of remaining executable cutting cycles for the current consumable part combination.

[0149] The mathematical computational model of the max-min normalization algorithm satisfies the following relationship:

[0150] ;

[0151] In the formula, These are the feature variable values ​​after minimax normalization. The input values ​​are the original feature variables. The minimum value of this feature variable obtained from the offline historical dataset; This represents the maximum value of the feature variable in this dimension obtained from statistics in the offline historical dataset.

[0152] S603, the long short-term memory neural network model relies on historical real-world datasets to optimize network node parameters during the offline phase. The external server obtains the multi-dimensional feature vector sequence extracted from the entire process of the plasma cutter from the installation of new consumable parts to the occurrence of cutting arc failure as the input of training samples. The external server uses the actual remaining number of cutting cycles experienced by each sampling node until the final scrapping as the supervision label of the corresponding sample. The external server uses mean squared error as the loss function and combines it with the adaptive moment estimation optimization algorithm to calculate the parameter gradient of the entire network architecture.

[0153] The external server iteratively updates the weight matrix and bias vector of the forget gate, input gate and output gate inside the hidden layer through the backpropagation mechanism until the loss function converges to the preset lower limit. In this embodiment, the preset lower limit ranges from 0.001 to 0.005. The external server sends the network weight parameters that have been trained offline through the industrial communication bus. The prediction module receives the network weight parameters and loads and statically fixes them in the underlying microcontroller.

[0154] The calculation model for the mean squared error loss function satisfies the following relationship:

[0155] ;

[0156] In the formula, This represents the mean squared error loss value output during model training. This represents the total number of samples included in a single network training batch. This is the index number of the sample sequence; For the first Supervision labels for the actual number of failure cycles corresponding to each sample; For long-term memory neural networks to the first The number of predicted failure cycles for each sample output, and the specific programming implementation of the data-driven model offline training on the underlying hardware, can be conventionally deployed using an open-source framework in this embodiment, which is a well-known technology in the field.

[0157] In this embodiment, step S7 provided by the present invention may include the following steps in specific implementation:

[0158] S701, the output module reads the total number of remaining executable cutting cycles of the current consumable parts combination output by the prediction module and compares it with the safety warning threshold set internally by the system. The total number of remaining executable cutting cycles of the current consumable parts combination is the smaller value between the number of remaining executable cutting cycles of the electrode and the number of remaining executable cutting cycles of the nozzle. The safety warning threshold is configured to 50 times according to the production line's part change and material preparation time requirements.

[0159] When the output module determines that the remaining number of cuts is lower than the set safety warning threshold, it pushes message data to the register address of the human-machine interface through the industrial communication interface to drive the display screen to render the consumable parts replacement prompt interface. The output module continuously monitors the predicted value. When it determines that the total number of remaining executable cutting cycles has dropped to the set forced shutdown limit threshold, which is configured based on the short-circuit impact resistance of the plasma cutting machine's main power supply and the workpiece scrap cost requirements, and the value range is set to 0 to 5 times, the output module activates the interrupt trigger logic in the hardware control circuit when it determines that the limit threshold condition is met.

[0160] The output module sends a forced shutdown level signal to the device's main control bus through a programmable logic pin and uses this level signal to intercept the next received cutting start command. The output module encapsulates all input feature vectors and accumulated loss status corresponding to the shutdown moment and writes them into the log area of ​​the system's non-volatile memory to form a closed-loop historical traceability sample record.

[0161] In this embodiment, step S8 provided by the present invention may include the following steps in specific implementation:

[0162] S801, the update module monitors the consumable replacement and reset commands issued by the human-machine interface in real time and determines that the current consumable usage cycle has ended when it receives the command. When the usage cycle has triggered an arc failure, a forced shutdown limit threshold, or when the operator enters the disassembly, inspection and scrapping confirmation information, the update module reads the total number of cycles actually executed within the usage cycle recorded in the system log area as the actual total number of failure cycles.

[0163] When the usage cycle is a preventative early replacement and has not met the scrap confirmation conditions, the update module marks the cycle as an incompletely failed sample, which is only used for historical archiving or auxiliary updates with confidence weights. The update module simultaneously extracts the basic integral term, electrode basic loss term before weighting, nozzle basic loss term before weighting, cross-stage splash penalty trigger state, voltage high-frequency envelope variance, and the time series of the mean absolute deviation of protective gas back pressure corresponding to each cutting cycle in the cycle.

[0164] The update module extracts the theoretical limit loss constants of electrodes and nozzles at the moment of physical scrapping, which are pre-calibrated and stored in the system through historical destructive tests, and constructs a global error objective function in combination with the aforementioned sequence data. This global error objective function quantifies the numerical deviation between the theoretical cumulative total loss calculated by the system using the current weight coefficients and the actual physical state of scrapping limit.

[0165] The mathematical model for calculating the global error objective function satisfies the following relationship:

[0166] ;

[0167] In the formula, The objective function for the three-dimensional global error after introducing cross-stage penalty parameters; The nozzle cross-excitation weighting coefficient; These are the electrode cross-excitation weighting coefficients; This refers to the cross-stage splash penalty weighting coefficient; This is the theoretical limiting loss constant for the electrode; This represents the total number of failure loops actually executed. The index number is the number of iterations. For the first three-dimensional parameter Single-cycle loss of electrode in each cycle; This represents the theoretical limiting loss constant of the nozzle. For the first three-dimensional parameter The nozzle loses power per cycle during each cycle.

[0168] S802, the update module calls the gradient descent algorithm based on the constructed global error objective function to perform the reverse closed-loop iterative optimization operation of the weight coefficients, and performs boundary projection on the nozzle cross-excitation weight coefficient, electrode cross-excitation weight coefficient and cross-stage splash penalty weight coefficient after each iteration, so that the three weight coefficients are kept within the preset physical effective range. The update module solves the first-order partial derivatives of the nozzle and electrode cross-excitation weight coefficient and the cross-stage splash penalty weight coefficient respectively to generate the corresponding three-dimensional gradient direction vector.

[0169] Before subtracting the product of the learning rate and the corresponding partial derivative from each weight coefficient, the update module extracts the absolute value of the partial derivative and compares it with a preset gradient clipping threshold. When the absolute value of the partial derivative exceeds the gradient clipping threshold, the partial derivative is forcibly limited to the upper limit of the threshold, and the iteration generation operation of the next generation of weight coefficient values ​​is performed. In this embodiment, the preset gradient clipping threshold ranges from 5 to 15. The update module compares the difference of the objective function calculated from two consecutive iterations with an internally fixed convergence tolerance threshold. When the absolute value of the difference is less than or equal to the convergence tolerance threshold or the current iteration count reaches the preset maximum iteration limit, the iteration operation is forcibly terminated and the weight coefficient value is output. In this embodiment, the internally fixed convergence tolerance threshold ranges from 1.0 × 10⁻⁶. -4 Up to 5.0×10 -4 .

[0170] The iterative update calculation model for the weight coefficients satisfies the following relationship:

[0171] ;

[0172] ;

[0173] ;

[0174] In the formula, To update the first generation of global error Secondary nozzle cross-excitation weighting coefficient; For the first Input values ​​for the nozzle cross-excitation weighting coefficients in the next iteration; The gradient descent learning rate constant is a preset value. In this embodiment, the value of the gradient descent learning rate constant ranges from 0.0001 to 0.001. The mathematical symbol for partial derivatives; The objective function is the three-dimensional global error. This represents the partial derivative of the global objective function with respect to the weight coefficient. To update the first generation of global error Secondary electrode cross-excitation weighting coefficient; For the first Input values ​​for electrode cross-excitation weighting coefficients in the next iteration; This represents the partial derivative of the global objective function with respect to the weight coefficient. For the first The cross-stage splash penalty weight coefficient generated by subgradient iteration; For the first Input values ​​for the cross-stage splash penalty weight coefficients in the next iteration; This represents the partial derivative of the global objective function with respect to the penalty weight coefficients. In this embodiment, a conventional numerical computing library can be used for engineering deployment, which is a well-known technology in the field.

[0175] S803, the update module performs a numerical validity check on the final three weight coefficient values ​​of the optimization output. The check includes that the values ​​are not empty, there is no overflow abnormality, there is no non-numerical abnormality, and they are within the preset physical valid range. After the check passes, the values ​​are overwritten to the specified low-level parameter configuration sector in the non-volatile memory. The update module uses the written weight coefficients to completely replace the old parameter matrix used in the previous life cycle. The update module simultaneously sends a hardware-level reset level pulse to the integration module.

[0176] The update module drives the integration module to perform a physical zeroing operation on the historical total wear variables of the electrodes and the historical total wear variables of the nozzle currently residing in the static random access memory. The update module erases the dynamic running data pointer of the previous cycle in the system log area and reinitializes the system's cache queue. The above hardware control operations enable the absolute time axis of the data link of the underlying controller and the physical state of consumables to return to the zero state synchronously and directly trigger a new round of continuous monitoring process of the life cycle.

[0177] Application Examples:

[0178] Please see the appendix Figure 7 - Appendix Figure 9 To better understand the technical solution of the present invention, the following is an application example of the continuous batch cutting of low carbon steel plates by a CNC plasma cutting machine in a heavy machinery manufacturing plant.

[0179] The acquisition module synchronously extracts the transient cutting voltage analog signal and cutting current analog signal of the inverter power supply. The acquisition module also acquires the real-time back pressure and ion gas flow rate output by the protective gas proportional valve in the auxiliary gas circuit at the same frequency. The multi-dimensional operating parameters are filled into the ring data buffer in the system's internal memory.

[0180] The preprocessing module calls the internally configured band-stop notch filter to filter high-frequency pulse interference generated by the switching power supply. The preprocessing module uses a low-pass filter to extract smooth electrical signals. The preprocessing module shifts the protective gas back pressure sequence and the ion gas flow rate sequence backward by a set number of discrete points according to the group delay constant of the filter. The preprocessing module completes the alignment of cross-physical field signals in the same absolute time coordinate system.

[0181] The identification module reads the first derivative of the smooth electrical signal and the back pressure deviation value. When the arc is established and the airflow reaches a balanced state, the identification module defines the corresponding time interval as the arc initiation segment. When the system continues to perform cutting operations, the identification module defines the corresponding time interval as the steady state segment. When the identification module receives a stop command and the cutting current decays to zero, the identification module defines the corresponding time interval as the arc termination segment.

[0182] The feature module extracts the variance of the high-frequency voltage envelope within the steady-state interval, calculates the mean absolute deviation of the protective gas back pressure within the steady-state interval, and generates a cross-coupling matrix using the extracted variance and mean. The cross-coupling matrix includes the weight values ​​of the accelerated effect of the mapped nozzle blockage state on electrode burn-off and the weight values ​​of the destructive effect of the mapped electrode wandering state on nozzle orifice expansion.

[0183] The integration module extracts current and voltage values ​​during the arc initiation phase and performs integration to generate transient thermal erosion loss variables. During the steady-state phase, the integration module extracts arc power and fluid scouring rate and performs integration with the cross-coupling matrix to generate cumulative loss variables. During the arc termination phase, when the variance of the high-frequency voltage envelope extracted during the steady-state phase exceeds the set wander limit threshold, the integration module activates the nonlinear penalty model to perform integration to generate splash additional loss variables. The integration module summarizes the loss variables generated in the above stages and outputs them as the total loss value for a single cutting cycle.

[0184] The prediction module adds the total loss value calculated at each stage of a single cutting cycle to the historical total loss variable of consumable parts recorded in non-volatile memory. The prediction module inputs the multi-dimensional feature sequence into the underlying long short-term memory neural network model to perform inference calculations and output the remaining number of cuts that can be performed on the current electrode and nozzle. When the prediction module determines that the predicted remaining number of cuts has decreased to the internally set forced shutdown limit threshold, it outputs a shutdown command to the main controller of the equipment.

[0185] When the update module confirms that a consumable part has been physically replaced, it reads the total number of actual failure cycles recorded in the previous life cycle. When there is a difference between the theoretical predicted life cycle value and the actual failure life cycle value, the update module constructs an objective function guided by minimizing the global error. The update module calls the gradient descent algorithm to perform the inverse operation of the matrix parameters to generate new cross-coupling weights and cross-stage splash penalty weights. The update module overwrites the verified new weight parameters into the underlying configuration sector as the basic variables for the system to start the next life cycle prediction operation.

[0186] Multi-cycle comparative tests were performed on the same plasma cutting machine. The test process used multiple sets of new electrodes with the same specifications and matching nozzles. The test group configuration adopted a first comparative strategy of forced replacement using fixed historical average number of cuts. The test group configuration adopted a second comparative strategy of monitoring the average change trend of a single cutting voltage to evaluate the life. The test group configuration adopted the multi-physics field feature fusion and cross-coupling evaluation and cross-stage spatter penalty long short-term memory neural network prediction strategy proposed in this invention.

[0187] Experimental test results show that the first comparison strategy, which does not take into account the actual dynamic operating parameters, produces premature failure underreporting. Experimental test results show that the second comparison strategy produces excessive life prediction deviation when encountering frequent switching between arc start-up and arc end-up conditions. Experimental test results show that the cross-coupled feature matrix calculation logic and arc end-up splash penalty mechanism operation logic of the strategy of this invention make the theoretical predicted scrap cycle number output by the model close to the actual cycle number when consumable parts suffer real physical damage. The strategy of this invention reduces the frequency of arc interruption failure of equipment and increases the extreme service life of a single set of consumable parts.

[0188] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A plasma cutting consumable parts life prediction system, characterized in that, include: The sensing module is used to acquire the raw analog signals of the multiphysics field; The acquisition module is used to convert the original analog signal into digital real-time operating parameters; A processing unit, connected to the acquisition module, integrates the following: The preprocessing module is used to filter the real-time operating parameters and perform a first-in-first-out queue delay alignment operation on the fluid dynamics signal; The identification module is used to perform dynamic time slicing based on working condition characteristics for a single work cycle; The feature module is used to extract physical degradation features and generate a cross-coupling coefficient matrix; The integration module is used to perform segmented loss accumulation and penalty feature calculation, and to summarize and generate the total loss. The prediction module is used to call the neural network model to perform remaining lifetime inference; The output module is used to output the predicted remaining useful life data; The update module is used to perform reverse iterative optimization on the nozzle cross-excitation weight coefficient, the electrode cross-excitation weight coefficient, and the cross-stage splash penalty weight coefficient.

2. The plasma cutting consumable parts life prediction system according to claim 1, characterized in that, The preprocessing module constructs a cascaded digital signal processing link in the low-frequency timing identification channel to perform traversal operations on the electrical sequence, extracts the switching frequency of the inverter main circuit as the notch center frequency to construct a band-stop notch filter, and connects a low-pass digital filter in series at the back end to output a smooth electrical feature sequence; at the same time, it constructs a high-frequency feature retention channel to output a high-frequency voltage feature sequence.

3. The plasma cutting consumable parts life prediction system according to claim 2, characterized in that, The identification module reads the real-time back pressure data of the protective gas after alignment and calculates the back pressure deviation of the protective gas. At the same time, it reads the first derivative of the filtered voltage, compares the absolute value of the first derivative of the filtered voltage with the preset zero bias threshold, confirms that the moment when the plasma arc and the protective gas flow field enter the thermodynamic equilibrium state is recorded as the steady-state boundary moment, and dynamically divides a single cutting cycle into the arc initiation segment, the steady-state segment, and the arc termination segment on the time axis.

4. The plasma cutting consumable parts life prediction system according to claim 3, characterized in that, The feature module generates a high-frequency voltage envelope sequence within the steady-state interval using a digital analytical signal construction method based on Hilbert transform. It calculates the mathematical variance of the high-frequency voltage envelope sequence as an electrode degradation feature and calculates the mean absolute deviation of the protective gas back pressure as a nozzle degradation feature. By introducing a linear weighting mechanism, it calculates the unidirectional excitation factor that aggravates the electrode arc migration caused by nozzle geometric anomalies and the unidirectional excitation factor that aggravates the burn-off effect of electrode emission anomalies on the nozzle flow channel, thereby generating the cross-coupling coefficient matrix.

5. A method for predicting the lifespan of consumable parts cut by plasma cutting, characterized in that, Using the plasma-cut consumable parts life prediction system according to any one of claims 1-4, the method includes the following steps: The acquisition module obtains digital real-time operating parameters, including transient cutting voltage, cutting current, real-time back pressure of protective gas, and ion gas flow rate. The preprocessing module performs multi-level feature-preserving filtering on the real-time running parameters and executes a time-delay alignment operation; The identification module combines control commands to perform dynamic timing boundary identification for a single work cycle, dynamically dividing the time axis into the arc initiation segment, steady-state segment, and arc termination segment. The feature module extracts electrode degradation features and nozzle degradation features, and calculates and generates a cross-coupling coefficient matrix. The integration module performs loss infinitesimal integral calculations according to the divided time intervals; The integration module summarizes the integral values ​​of each segment to generate the total loss, and the prediction module performs inference and prediction calculations of the remaining service life through a long short-term memory neural network model. The output module outputs a consumable parts replacement prompt when the remaining service life is lower than the safety warning threshold, and outputs a shutdown protection command when the remaining service life is lower than the forced shutdown limit threshold. The update module uses the gradient descent algorithm to perform closed-loop iterative optimization of the nozzle cross-excitation weight coefficient, the electrode cross-excitation weight coefficient, and the cross-stage splash penalty weight coefficient.

6. The method for predicting the lifespan of consumable parts cut by plasma cutting according to claim 5, characterized in that, The steps by which the acquisition module obtains digitized real-time operating parameters include: The acquisition module synchronously discretizes the analog electrical signals of transient cutting voltage and cutting current, and uses the high-frequency sampling clock for acquiring electrical parameters to convert the analog signals related to fluid dynamics. It forces the fluid discrete sequence and the electrical discrete sequence to have the same time axis coordinate density, and pushes the data into the ring data buffer for continuous storage after aligning the data according to the order of generation time.

7. The method for predicting the lifespan of consumable parts cut by plasma cutting according to claim 6, characterized in that, The preprocessing module performs multi-level feature-preserving filtering on real-time operating parameters and executes time-delay alignment operations, including the following steps: The preprocessing module calculates the absolute time lag introduced by the digital filtering link in the data channel as the group delay parameter, and calculates the offset of discrete data points in combination with the high-frequency sampling frequency; a dedicated first-in-first-out buffer queue is planned to push the discrete fluid sequence into the buffer, and the offset of the discrete data points is used to perform a hysteresis locking operation on the dequeue pointer to force the fluid signal to generate a physical residence time and complete the time axis synchronization.

8. The method for predicting the lifespan of plasma-cut consumable parts according to claim 7, characterized in that, The steps of performing loss infinitesimal integral calculations by the integration module according to the divided time intervals include: The amplitude square and the absolute value of the first derivative of each data point in the discrete sequence of cutting current are extracted, multiplied by the corresponding constant coefficients, and then summed to generate a transient erosion micro-element function to calculate the transient impact loss of the arc segment; the unidirectional excitation factor is extracted and multiplied by the electrode basic micro-element loss rate and the nozzle basic micro-element loss rate to generate a steady-state integral function to calculate the coupling cumulative loss of the steady-state segment; when the variance of the high-frequency envelope of the voltage exceeds the wander limit threshold, the nonlinear penalty integral calculation logic is activated to calculate the additional splash loss.

9. The method for predicting the lifespan of plasma-cut consumable parts according to claim 8, characterized in that, The prediction module performs the following steps to calculate the remaining useful life using a long short-term memory neural network model: The current cumulative electrode wear, current cumulative nozzle wear, high-frequency voltage envelope variance, and mean absolute deviation of protective gas back pressure from the current cutting cycle and several previous consecutive cutting cycles are combined in a fixed order to construct a two-dimensional feature matrix sequence. After linearly mapping the values ​​using the max-min normalization algorithm, the sequence is input into the long short-term memory neural network model. The output results are obtained from the linear fully connected layer, and a non-negative truncation operation is performed to output the remaining number of executable cutting cycles for the electrode and nozzle respectively. The smaller of the two values ​​is taken as the total number of remaining executable cutting cycles for the current consumable part combination.

10. The method for predicting the lifespan of plasma-cut consumable parts according to claim 9, characterized in that, The update module employs a gradient descent algorithm to perform closed-loop iterative optimization of the nozzle cross-excitation weight coefficient, electrode cross-excitation weight coefficient, and cross-stage splash penalty weight coefficient. The steps include: Extract the theoretical limit loss constant from the calibrated storage and construct a global error objective function to quantify the deviation between the theoretical cumulative total loss and the actual scrap limit physical state by combining historical cyclic data; solve the first-order partial derivatives of the nozzle cross-excitation weight coefficient, electrode cross-excitation weight coefficient, and cross-stage splash penalty weight coefficient respectively to generate a three-dimensional gradient direction vector; after performing gradient absolute value comparison and upper limit limiting, iteratively generate a new generation of nozzle cross-excitation weight coefficient, electrode cross-excitation weight coefficient, and cross-stage splash penalty weight coefficient, and overwrite them into the non-volatile memory to replace the old parameter matrix.

Images