Intelligent prediction method and system for hall thruster propellant residual

Abstract

The method acquires operation data of the target Hall thruster, extracts a shutdown recovery section corresponding to each propelling action, and constructs a recovery dynamics atlas; according to the recovery dynamics atlas and a reference recovery dynamics atlas under a corresponding working condition, a margin-sensitive characteristic quantity is extracted; the margin-sensitive characteristic quantity is sequentially arranged in time order to form a time sequence sample sequence, and is input into a deep learning prediction model to obtain a current propellant margin prediction result and a future propellant margin prediction result. The scheme uses the recovery process after the propelling action to represent the propellant margin change, and reduces the influence of transient disturbance and working condition difference in the propelling process on the margin prediction.

Description

Technical Field

[0001] This invention relates to the field of Hall thruster propellant balance prediction technology, specifically to an intelligent prediction method and system for Hall thruster propellant balance. Background Technology

[0002] Hall thrusters are typical space electric propulsion devices. Due to their high specific impulse, long service life, and low propellant consumption, they have gradually evolved from early engineering verification into an important propulsion means for satellite orbit maintenance, orbit raising, and deep space exploration. NASA data shows that Hall thrusters such as SPT-100 have already had many flight applications; ESA's SMART-1 mission verified the long-term on-orbit operation capability of Hall thrusters as the main propulsion system, indicating that this technology has entered the stage of engineering and large-scale application.

[0003] Current methods for estimating propellant reserves in spacecraft primarily employ bookkeeping and pressure-volume-temperature (PCT) correlation methods. The former accumulates errors in flow rate estimation, while the latter relies on tank pressure and temperature measurements and is susceptible to thermal state changes and mission disturbances. For Hall thrusters, which experience frequent start-stop cycles, operational mode switching, and thermal coupling recovery processes, these methods struggle to simultaneously guarantee the accuracy and stability of both current reserve estimates and future reserve trend predictions. Therefore, intelligent prediction methods and systems for Hall thruster propellant reserves are needed to address these issues. Summary of the Invention

[0004] (a) Technical problem to be solved: In view of the shortcomings of the prior art, the present invention provides an intelligent prediction method and system for propellant surplus of Hall thrusters, which solves the above-mentioned problems.

[0005] (II) Technical Solution: To achieve the above objectives, the present invention provides the following technical solution: an intelligent prediction method for propellant balance in Hall thrusters, comprising: S1. Obtain the operating data of the target Hall thruster, wherein the operating data includes at least tank pressure, tank temperature, valve status, discharge voltage, discharge current, thrust command and operating mode; S2. Extract the shutdown recovery segment for each propulsion action, and construct a recovery dynamics map based on each shutdown recovery segment; S3. Based on the recovery dynamics maps described above and the baseline recovery dynamics maps under the corresponding working conditions, extract the residual sensitive feature quantities; S4. The remaining sensitive features of each propulsion action are arranged into a time series sample sequence in chronological order and input into the deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period.

[0006] Furthermore, the extraction of the shutdown recovery segment includes: Based on the valve status changing from open to closed and the discharge current dropping below the preset cutoff threshold, determine the shut-off starting point corresponding to each propulsion action; Data on the changes in tank pressure and tank temperature over time are extracted from each of the aforementioned shutdown points; If the rate of change of tank pressure is less than a first threshold and the rate of change of tank temperature is less than a second threshold within a continuous preset time period, then the start time of the continuous preset time period is taken as the end time of recovery. The data between each shutdown start point and the corresponding recovery end time is defined as the shutdown recovery segment.

[0007] Furthermore, the construction of the recovery dynamics map based on each of the shutdown recovery segments includes: The tank pressure and tank temperature data in each of the shutdown recovery sections are normalized to obtain normalized recovery data; Based on the normalized recovery data, a time-domain recovery subplot, a pressure-temperature phase diagram subplot, and a recovery rate subplot are constructed; wherein, the time-domain recovery subplot includes a normalized pressure recovery curve and a normalized temperature recovery curve, the pressure-temperature phase diagram subplot is a coupled trajectory formed by pressure and temperature in time sequence, and the recovery rate subplot includes a pressure recovery rate curve and a temperature recovery rate curve. The time-domain recovery sub-map, the pressure-temperature phase diagram sub-map, and the recovery rate sub-map are combined to obtain the recovery kinetics map.

[0008] Furthermore, the normalization processing of the tank pressure and tank temperature data within each of the shutdown and recovery sections includes: Time normalization processing is performed on the time data corresponding to each sampling time within each of the aforementioned shutdown recovery segments; Based on the tank pressure corresponding to the shutdown start point and the tank pressure corresponding to the recovery end point in each shutdown recovery segment, the tank pressure data in each shutdown recovery segment are normalized to obtain a normalized pressure recovery curve. Based on the tank temperature corresponding to the shutdown start point and the tank temperature corresponding to the recovery end time in each shutdown recovery segment, the tank temperature data in each shutdown recovery segment are normalized to obtain a normalized temperature recovery curve. Based on the normalized pressure recovery curve and the normalized temperature recovery curve, the pressure recovery rate curve and the temperature recovery rate curve are obtained respectively.

[0009] Furthermore, the baseline recovery kinetics map under the corresponding operating condition includes: Historical propulsion actions are classified according to at least one of the following: thrust command level, operating mode, average discharge current range, average discharge voltage range, and propulsion duration range. Time alignment and statistical processing are performed on multiple historical recovery kinetic maps belonging to the same working condition category to obtain the baseline recovery kinetic map under the corresponding working condition; The reference recovery kinetic map includes at least a reference time domain recovery sub-map, a reference pressure-temperature phase map sub-map, and a reference recovery rate sub-map. The reference time domain recovery sub-map includes a reference pressure recovery curve and a reference temperature recovery curve, and the reference recovery rate sub-map includes a reference pressure recovery rate curve and a reference temperature recovery rate curve.

[0010] Furthermore, S3 includes: The deviation information is obtained by comparing the recovery dynamics map corresponding to the current propulsion action with the baseline recovery dynamics map under the corresponding working condition. The deviation information includes at least the difference information between the current time domain recovery subgraph and the reference time domain recovery subgraph, the geometric deviation information between the current pressure-temperature phase diagram subgraph and the reference pressure-temperature phase diagram subgraph, and the difference information between the current recovery rate subgraph and the reference recovery rate subgraph. Based on the recovery dynamics map corresponding to the current propulsion action and the deviation information, the residual sensitive feature quantity is extracted.

[0011] Furthermore, the margin-sensitive feature quantity includes at least one or more of the following: Pressure half-recovery time, temperature half-recovery time, hysteresis between pressure recovery peak and temperature recovery peak, area enclosed by pressure-temperature phase diagram, centroid coordinates of pressure-temperature phase diagram, principal axis direction angle of pressure-temperature phase diagram, integral difference between current normalized pressure recovery curve and reference pressure recovery curve, integral difference between current normalized temperature recovery curve and reference temperature recovery curve, difference between current pressure recovery rate curve and reference pressure recovery rate curve, and difference between current temperature recovery rate curve and reference temperature recovery rate curve.

[0012] Furthermore, S4 includes: The remaining sensitive features of each propulsion action and the propulsion condition parameters corresponding to each propulsion action are arranged in chronological order to form a time-series sample sequence. The time-series sample sequence is input into a deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period; The propulsion condition parameters include at least one or more of the following: propulsion duration, average discharge current, average discharge voltage, thrust command level, and operating mode. The deep learning prediction model is a long short-term memory network, a gated recurrent unit network, or a Transformer network.

[0013] The present invention also provides an intelligent prediction system for propellant margin in Hall thrusters, comprising: The data acquisition module is used to acquire the operating data of the target Hall thruster. The operating data includes at least tank pressure, tank temperature, valve status, discharge voltage, discharge current, thrust command, and operating mode. The graph construction module is used to extract the shutdown recovery segment for each propulsion action and construct a recovery dynamics graph based on each shutdown recovery segment; The feature extraction module is used to extract residual sensitive features based on the recovery kinetic maps and the benchmark recovery kinetic maps under the corresponding working conditions. The prediction module is used to assemble the remaining sensitive features of each propulsion action into a time-series sample sequence in chronological order, and input it into the deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period.

[0014] (III) Beneficial Effects: Compared with the prior art, the present invention provides an intelligent prediction method and system for propellant margin of Hall thrusters, which has the following beneficial effects: 1. The intelligent prediction method and system for propellant balance of Hall thrusters extracts the shut-off recovery segment for each thrusting action and constructs a recovery dynamics map based on the shut-off recovery segment. The pressure recovery behavior, temperature recovery behavior and their coupling change relationship after the thrusting process are used as indirect characterization objects of propellant balance. This avoids the fluctuation interference and cumulative error problems caused by the excessive reliance of existing methods on instantaneous pressure value, instantaneous temperature value or cumulative flow value during the thrusting process, and improves the stability and physical consistency of propellant balance characterization.

[0015] 2. The intelligent prediction method and system for propellant balance of Hall thrusters extracts balance-sensitive features based on the recovery kinetics map and the baseline recovery kinetics map under the corresponding operating conditions. It compares the recovery behavior after the current thrust action with the baseline recovery behavior under the same operating conditions to obtain deviation information that reflects the change in balance state. This reduces the confusion caused by different thrust commands, different operating modes and different thrust durations on the balance estimation results, solves the problem that the existing pressure-temperature estimation method is susceptible to changes in thermal state and disturbances in the mission process, and improves the accuracy of propellant balance identification under complex operating conditions. Attached Figure Description

[0016] Figure 1 A schematic diagram illustrating the steps of the intelligent prediction method for propellant margin in Hall thrusters provided by the present invention; Figure 2 A flowchart illustrating the intelligent prediction method for propellant margin in Hall thrusters provided by this invention. Figure 3 This is a schematic diagram of the intelligent prediction system for propellant margin of Hall thrusters provided by the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Please see Figure 1-2 , Figure 1 A flowchart illustrating the intelligent prediction method for propellant margin in Hall thrusters provided by this invention. Figure 2 This is a flowchart illustrating the intelligent prediction method for propellant margin in Hall thrusters provided by the present invention; this application provides an intelligent prediction method for propellant margin in Hall thrusters, comprising: S1. Obtain the operating data of the target Hall thruster, wherein the operating data includes at least tank pressure, tank temperature, valve status, discharge voltage, discharge current, thrust command and operating mode; Specifically, tank pressure is collected via a pressure sensor, tank temperature via a temperature sensor, valve status is determined via valve drive feedback signals, discharge voltage and discharge current are obtained via power telemetry signals, and thrust commands and operating modes are obtained via the propulsion controller. A unified time stamp is added to each data point, and the data is arranged in chronological order to form an operational data sequence. When sampling times from different data sources are inconsistent, time alignment is performed using a unified time base to obtain the correspondence between pressure, temperature, valve status, discharge voltage, discharge current, thrust commands, and operating modes at each sampling time. S2. Extract the shutdown recovery segment for each propulsion action, and construct a recovery dynamics map based on each shutdown recovery segment; Specifically, the end time of each propulsion action is first identified based on the operational data. Then, the corresponding shut-off recovery segment is extracted from the pressure recovery process and temperature recovery process after each propulsion action. Next, the pressure and temperature data in each shut-off recovery segment are normalized, and time-domain recovery sub-graphs, pressure-temperature phase diagram sub-graphs, and recovery rate sub-graphs are constructed respectively. Finally, the time-domain recovery sub-graphs, pressure-temperature phase diagram sub-graphs, and recovery rate sub-graphs corresponding to the same propulsion action are combined into a unified map record in a fixed layout to form the recovery kinetic map corresponding to that propulsion action.

[0020] In this model, the three sub-plots corresponding to the same propulsion action are associated using the same propulsion action number, and employ the same normalized time range and the same number of sampling points. When combining the plots, the time-domain recovery sub-plot is placed on the left, the pressure-temperature phase diagram sub-plot in the middle, and the recovery rate sub-plot on the right; simultaneously, the numerical sequences corresponding to the three sub-plots are saved as direct input for subsequent extraction of residual sensitive features. That is, the recovered kinetics plot includes not only the visual layout of the three sub-plots, but also the plots related to the three sub-plots. Figure 1 A corresponding numerical dataset.

[0021] Furthermore, in one embodiment of the present invention, the extraction of the shutdown recovery segment includes: Based on the valve status changing from open to closed and the discharge current dropping below the preset cutoff threshold, determine the shut-off starting point corresponding to each propulsion action; Data on the changes in tank pressure and tank temperature over time are extracted from each of the aforementioned shutdown points; If the rate of change of tank pressure is less than a first threshold and the rate of change of tank temperature is less than a second threshold within a continuous preset time period, then the start time of the continuous preset time period is taken as the end time of recovery. The data between each shutdown start point and the corresponding recovery end time is defined as the shutdown recovery segment.

[0022] Specifically, let the sampling period be... First, find the sampling point in the operating data where the valve state changes from open to closed, and record it as a candidate shut-off point; then, starting from the candidate shut-off point, detect the discharge current backwards. If the discharge current is continuous... The discharge current at each sampling point is less than the preset cutoff threshold. The first sampling point that meets the condition is recorded as the turn-off start point. .

[0023] from Begin by extracting the tank pressure sequence in chronological order. and tank temperature sequence To reduce the impact of single-point noise, the pressure and temperature sequences are first processed using a moving average, with a moving window length of 3 to 5 sampling points. Then, the pressure and temperature change rates are calculated using the difference between adjacent sampling points, i.e.: and .

[0024] from Search backwards, when consecutive Each sampling point simultaneously satisfies and At that time, this The first sampling time among the sampling points is recorded as the recovery end time. .in, The first threshold, The second threshold is the continuous preset duration equal to... .

[0025] Ultimately, arrive The pressure, temperature, and corresponding time data are extracted and used as the shutdown recovery segment for each propulsion action. If the preset maximum duration is still not met to end the recovery, the propulsion action is marked as an invalid sample and removed, and will not proceed to the subsequent map construction process.

[0026] Furthermore, in one embodiment of the present invention, the construction of a recovery dynamics map based on each of the shutdown recovery segments includes: The tank pressure and tank temperature data in each of the shutdown recovery sections are normalized to obtain normalized recovery data; Based on the normalized recovery data, a time-domain recovery subplot, a pressure-temperature phase diagram subplot, and a recovery rate subplot are constructed; wherein, the time-domain recovery subplot includes a normalized pressure recovery curve and a normalized temperature recovery curve, the pressure-temperature phase diagram subplot is a coupled trajectory formed by pressure and temperature in time sequence, and the recovery rate subplot includes a pressure recovery rate curve and a temperature recovery rate curve. The time-domain recovery sub-map, the pressure-temperature phase diagram sub-map, and the recovery rate sub-map are combined to obtain the recovery kinetics map.

[0027] Specifically, after normalizing each shutdown recovery segment, the normalized time series, normalized pressure recovery series, and normalized temperature recovery series are resampled to a fixed number of points. For example, 64 points, 128 points, or 256 points are used to ensure the same data length across different propulsion actions. The resampling method employs linear interpolation.

[0028] After resampling, construct the three types of subgraphs as follows: The time-domain recovery subplot is generated by plotting two recovery curves: the resampled normalized time series on the horizontal axis and the resampled normalized pressure and temperature recovery series on the vertical axes. This subplot directly reflects the changes in pressure and temperature recovery over normalized time.

[0029] The pressure-temperature phase diagram subplot is obtained by connecting the sampling points in the sampling order, with the resampled normalized pressure recovery sequence as the horizontal axis and the resampled normalized temperature recovery sequence as the vertical axis, forming a coupled trajectory with time sequence. This subplot reflects the correspondence between pressure recovery and temperature recovery.

[0030] The recovery rate subplot is generated by first calculating the rate of change between adjacent sampling points based on the resampled normalized pressure recovery sequence and the normalized temperature recovery sequence, respectively, to obtain the pressure recovery rate sequence and the temperature recovery rate sequence. Then, two rate curves are plotted with the normalized time series as the horizontal axis and the pressure recovery rate sequence and the temperature recovery rate sequence as the vertical axis, resulting in the recovery rate subplot. This subplot reflects the change in the recovery rate over time.

[0031] After the three types of subplots are generated, they are arranged in left, center, and right order on a fixed-size two-dimensional canvas: the left side is the time-domain recovery subplot, the center side is the pressure-temperature phase diagram subplot, and the right side is the recovery rate subplot. The canvas size, the coordinate axis range of the three types of subplots, and the subplot spacing are all set to be fixed.

[0032] While completing the graphic combination, the three sets of numerical data corresponding to the same propulsion action are saved in a unified format, including: the data corresponding to the time-domain reconstruction subgraph. The data corresponding to the pressure-temperature phase diagram sub-diagram is: The data corresponding to the recovery rate subplot is ,in, Finally, the combined image of the three sub-graphs and the three sets of numerical data are used as the output of the recovery dynamics map for this propulsion action.

[0033] Furthermore, in one embodiment of the present invention, the normalization processing of the tank pressure and tank temperature data within each of the shutdown recovery sections includes: Time normalization processing is performed on the time data corresponding to each sampling time within each of the aforementioned shutdown recovery segments; Based on the tank pressure corresponding to the shutdown start point and the tank pressure corresponding to the recovery end point in each shutdown recovery segment, the tank pressure data in each shutdown recovery segment are normalized to obtain a normalized pressure recovery curve. Based on the tank temperature corresponding to the shutdown start point and the tank temperature corresponding to the recovery end time in each shutdown recovery segment, the tank temperature data in each shutdown recovery segment are normalized to obtain a normalized temperature recovery curve. Based on the normalized pressure recovery curve and the normalized temperature recovery curve, the pressure recovery rate curve and the temperature recovery rate curve are obtained respectively.

[0034] Specifically, let the shutdown start point be... The recovery end time is The first section of the shutdown and recovery process Each sampling time is Then the normalized time is defined as: After this processing, the start time of any shutdown recovery segment is mapped to 0, and the end time is mapped to 1.

[0035] Let the tank pressure corresponding to the shutdown start point be... The tank pressure at the end of the recovery process is The tank pressure at any sampling moment within the shutdown-recovery phase is The normalized pressure recovery is defined as: .

[0036] Let the tank temperature corresponding to the shutdown start point be... The tank temperature at the end of the recovery process is , No. The tank temperature corresponding to each sampling point is: The normalized temperature recovery is defined as: .

[0037] when When the pressure difference is less than the preset pressure difference threshold, the shut-off recovery segment is marked as a sample with insufficient pressure recovery amplitude; when When the temperature difference is less than a preset threshold, the segment of the shutdown recovery is marked as a sample with insufficient temperature recovery amplitude. Samples with insufficient amplitude can be directly removed, or only another type of valid recovery curve can be retained for subsequent processing.

[0038] After obtaining the normalized pressure recovery curve and the normalized temperature recovery curve, the recovery rate is calculated using the difference between adjacent sampling points. If the normalized time interval after resampling is... , then the first The pressure recovery rate and temperature recovery rate corresponding to each resampling point are as follows: , ;in, The recovery rate of the first sampling point is the same as that of the second sampling point, or it can be calculated using the forward difference method.

[0039] After the above processing, the shutdown recovery segments corresponding to different propulsion actions are uniformly mapped to the same time range, the same recovery range, and the same number of sampling points, so that the recovery dynamics map can be constructed in a unified format.

[0040] S3. Based on the recovery dynamics maps described above and the baseline recovery dynamics maps under the corresponding working conditions, extract the residual sensitive feature quantities; Specifically, first, based on the thrust command level, operating mode, average discharge current, average discharge voltage, and propulsion duration corresponding to the current propulsion action, the operating condition category of the current propulsion action is determined; then, the benchmark recovery dynamics map corresponding to this operating condition category is retrieved from the benchmark recovery dynamics map library; next, the recovery dynamics map corresponding to the current propulsion action is compared item by item with the benchmark recovery dynamics map to obtain the time domain recovery deviation result, pressure-temperature coupling deviation result, and recovery rate deviation result; finally, based on the recovery morphology information of the current map itself and various deviation results, the margin sensitive feature quantity is extracted, and each feature quantity is arranged in a fixed order to form the feature vector corresponding to this propulsion action.

[0041] If the current propulsion action simultaneously meets the classification criteria for multiple operating conditions, the operating condition category with the smallest comprehensive distance from the current thrust command level, operating mode, average discharge current, average discharge voltage, and propulsion duration is selected as the matching result. If there is no perfectly matching operating condition category, the category with the largest number of samples among adjacent operating condition categories is selected as the alternative operating condition category. The comprehensive distance is calculated using the Euclidean distance after normalization of each operating condition parameter.

[0042] Furthermore, in one embodiment of the present invention, the baseline recovery kinetics spectrum under the corresponding operating condition includes: Historical propulsion actions are classified according to at least one of the following: thrust command level, operating mode, average discharge current range, average discharge voltage range, and propulsion duration range. Time alignment and statistical processing are performed on multiple historical recovery kinetic maps belonging to the same working condition category to obtain the baseline recovery kinetic map under the corresponding working condition; The reference recovery kinetic map includes at least a reference time domain recovery sub-map, a reference pressure-temperature phase map sub-map, and a reference recovery rate sub-map. The reference time domain recovery sub-map includes a reference pressure recovery curve and a reference temperature recovery curve, and the reference recovery rate sub-map includes a reference pressure recovery rate curve and a reference temperature recovery rate curve.

[0043] Specifically, historical propulsion actions are categorized into multiple operating condition classes based on operating parameters. Thrust command levels are classified by discrete command values; operating modes are classified by mode number; average discharge current and average discharge voltage are respectively grouped into preset intervals; and propulsion duration is grouped into preset duration intervals. For all historical propulsion actions within the same operating condition class, three sets of numerical data are read from the corresponding reconstructed dynamics graph, i.e. , and Since the previous step had already resampled the recovery process of each propulsion action to the same number of points, Therefore, different samples within the same working condition category are naturally aligned at the sampling point location, and no separate interpolation alignment is required.

[0044] Let the number of valid historical samples within the same working condition category be . Then for the first For each sampling point, the reference pressure recovery amount, reference temperature recovery amount, reference pressure recovery rate, and reference temperature recovery rate are respectively taken as: , , , ;in, Indicates the first A historical sample, .

[0045] based on Construct a baseline time-domain recovery subgraph; based on Connect the samples in the sampling order to form a sub-diagram of the reference pressure-temperature phase diagram; based on Construct a baseline recovery rate subgraph.

[0046] When the number of valid historical samples within the same working condition category is less than the preset minimum number of samples, the working condition category is merged with adjacent working condition categories and the statistics are recalculated, or the working condition category is marked as a missing category in the baseline map and is not included in subsequent feature extraction.

[0047] Furthermore, in one embodiment provided by the present invention, S3 includes: The deviation information is obtained by comparing the recovery dynamics map corresponding to the current propulsion action with the baseline recovery dynamics map under the corresponding working condition. The deviation information includes at least the difference information between the current time domain recovery subgraph and the reference time domain recovery subgraph, the geometric deviation information between the current pressure-temperature phase diagram subgraph and the reference pressure-temperature phase diagram subgraph, and the difference information between the current recovery rate subgraph and the reference recovery rate subgraph. Based on the recovery dynamics map corresponding to the current propulsion action and the deviation information, the residual sensitive feature quantity is extracted.

[0048] Specifically, for the time-domain recovery subplot, the current normalized pressure recovery curve is compared with the reference pressure recovery curve, and the current normalized temperature recovery curve is compared with the reference temperature recovery curve point by point according to the sampling point, to obtain: , .

[0049] For the recovery rate subplot, by comparing the current pressure recovery rate curve with the reference pressure recovery rate curve and the current temperature recovery rate curve with the reference temperature recovery rate curve point by point, we obtain: , .

[0050] For the pressure-temperature phase diagram sub-diagram, geometric deviation information is characterized by the following three types of quantities: The first category is the area deviation, which is the difference between the area of ​​the current pressure-temperature phase diagram and the area of ​​the reference pressure-temperature phase diagram; The second category is the centroid deviation, which is the coordinate difference between the centroid of the current pressure-temperature phase diagram and the centroid of the reference pressure-temperature phase diagram; The third category is the directional deviation, which is the difference between the principal axis direction angle of the current pressure-temperature phase diagram and the principal axis direction angle of the reference pressure-temperature phase diagram.

[0051] After the comparison is completed, , , , In addition, geometric deviation information is output as deviation information. Then, the current recovery morphological parameters are extracted from the current recovery dynamics map, and the relative deviation parameters are extracted from the deviation information, finally forming a set of residual sensitive features.

[0052] Furthermore, in one embodiment of the present invention, the margin-sensitive feature quantity includes at least one or more of the following: Pressure half-recovery time, temperature half-recovery time, hysteresis between pressure recovery peak and temperature recovery peak, area enclosed by pressure-temperature phase diagram, centroid coordinates of pressure-temperature phase diagram, principal axis direction angle of pressure-temperature phase diagram, integral difference between current normalized pressure recovery curve and reference pressure recovery curve, integral difference between current normalized temperature recovery curve and reference temperature recovery curve, difference between current pressure recovery rate curve and reference pressure recovery rate curve, and difference between current temperature recovery rate curve and reference temperature recovery rate curve.

[0053] Specifically, the pressure half-recovery time is taken as the normalized time value corresponding to the first time the current normalized pressure recovery curve reaches 0.5; if 0.5 falls between two adjacent sampling points, linear interpolation is used to calculate the corresponding time. The temperature half-recovery time is obtained from the normalized temperature recovery curve in the same way.

[0054] The hysteresis between the pressure recovery peak and the temperature recovery peak is taken as the difference between the normalized time corresponding to the maximum value of the pressure recovery rate curve and the normalized time corresponding to the maximum value of the temperature recovery rate curve. If the time corresponding to the peak value of the pressure recovery rate curve is... The peak time corresponding to the temperature recovery rate curve is The lag quantity is then taken as .

[0055] The area enclosed by the pressure-temperature phase diagram is calculated using the polygon area formula. Let the point sequence formed by the pressure-temperature phase diagram in the sampling order be... Then the area is taken as .

[0056] The centroid coordinates of the pressure-temperature phase diagram are taken as the average of the x and y coordinates of all sampling points, i.e. , .

[0057] The principal axis direction angle of the pressure-temperature phase diagram is obtained by finding the angle between the eigenvector corresponding to the largest eigenvalue and the horizontal axis through the covariance matrix of all points.

[0058] The integral difference between the current normalized pressure recovery curve and the reference pressure recovery curve is taken as... The discrete integral value over the entire sampling interval; the integral difference between the current normalized temperature recovery curve and the reference temperature recovery curve, taken as... The discrete integral value over the entire sampling interval. If the normalized time interval after resampling is... Then the two are respectively taken as , .

[0059] The differences between the current pressure recovery rate curve and the reference pressure recovery rate curve, and the differences between the current temperature recovery rate curve and the reference temperature recovery rate curve, are respectively calculated using point-by-point difference sequences. and This means that, before being used as model input, the two sets of difference sequences can be further compressed into statistics of fixed dimension, such as mean, maximum, minimum, and integral values.

[0060] Finally, all the above features are arranged in a fixed order to form a feature vector. For example, the inherent features of the current recovery kinetics map are arranged first, followed by the deviation features from the baseline recovery kinetics map; once the arrangement order is determined, it remains consistent during the training and inference phases. In this way, each advancement action corresponds to a margin-sensitive feature vector with a fixed dimension and clear meaning, which is then used to form a time-series sample sequence and input into the deep learning prediction model.

[0061] S4. The remaining sensitive features of each propulsion action are arranged into a time series sample sequence in chronological order and input into the deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period.

[0062] Specifically, for each propulsion action, the residual sensitive feature quantities corresponding to the propulsion action are first arranged in a fixed order to form a single-step feature vector; then, according to the order in which the propulsion actions occur, the single-step feature vectors corresponding to multiple consecutive propulsion actions are combined into a time-series sample sequence.

[0063] Let the first The single-step feature vector corresponding to the next propulsion action is Take continuous The single-step feature vectors corresponding to each propulsion action form an input sample, i.e. ;in, For the preset sequence length, Number the current propulsion action. If the number of valid propulsion actions before the current propulsion action is less than... Then, the sequence is padded to the length using one of the following methods: zero padding, repeating the first valid vector, or introducing a preset null value marker. .

[0064] Before inputting the time-series sample sequences into the deep learning prediction model, the features of each dimension are first numerically normalized. The normalization parameters are obtained statistically from the training set and remain consistent during the training and inference phases.

[0065] After receiving a time-series sample sequence, the deep learning prediction model outputs two results: first, a predicted value of the current propellant remaining quantity corresponding to the current propulsion action; and second, a sequence of predicted propellant remaining quantities within a preset future time period or a preset number of future propulsion actions. If the future prediction range is based on future... If the next propulsion action is set, the model output can be expressed as: ;in, This represents the current predicted propellant balance. Indicates the future number Predicted propellant remaining amount after each propulsion maneuver.

[0066] During model training, the known propellant balance values ​​corresponding to historical samples are used as supervision labels, and the mean squared error loss function is used to iteratively update the model parameters. After the model is deployed, the newly formed time series sample sequence is used as input, and the current propellant balance prediction result and the future propellant balance prediction result are directly output.

[0067] Furthermore, in one embodiment provided by the present invention, S4 includes: The remaining sensitive features of each propulsion action and the propulsion condition parameters corresponding to each propulsion action are arranged in chronological order to form a time-series sample sequence. The time-series sample sequence is input into a deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period; The propulsion condition parameters include at least one or more of the following: propulsion duration, average discharge current, average discharge voltage, thrust command level, and operating mode. The deep learning prediction model is a long short-term memory network, a gated recurrent unit network, or a Transformer network.

[0068] Specifically, for the first The next step involves extracting the residual sensitive feature vector. Then extract the propulsion condition parameter vector. Among them, the propulsion duration is the duration from the start of ignition to the start of shutdown of the propulsion action; the average discharge current is the average of all discharge current samples during the propulsion action; the average discharge voltage is the average of all discharge voltage samples during the propulsion action; the thrust command level is represented by discrete encoding; and the operating mode is represented by one-hot encoding or integer encoding.

[0069] After that and By assembling them in a fixed order, we obtain the [number]. The combined feature vector corresponding to each propulsion action: Then, following the ascending order of the propulsion action numbers, the consecutive... The time-series sample sequence is composed of a combination of feature vectors: .

[0070] When a deep learning prediction model uses a long short-term memory network or a gated recurrent unit network, the time series sample sequence is input into the network in the form of a tensor of sequence length × feature dimension. The recurrent unit reads the input at each time step in sequence, and finally takes the hidden state at the last time step or the hidden states at all time steps, maps them through a fully connected layer, and outputs the current propellant balance prediction value and the future propellant balance prediction value sequence.

[0071] When a deep learning prediction model uses a Transformer network, it first encodes the additional position of each time step in the time series sample sequence, then inputs it into the encoder structure for feature extraction, and finally obtains the current propellant balance prediction value and the future propellant balance prediction value sequence through the output layer.

[0072] During the training phase, multiple time-series sample sequences are generated from the historical propagation action sequences using a sliding window method; for example, taking the first... If the next advancement action is taken as the current moment, then the first advancement action is taken. Next to The combined feature vector of the propulsion action is used as input, with the first propulsion action as the second propulsion action as the third propulsion action. The actual propellant balance corresponding to the first propulsion action is used as the current label, with the first... Next to The actual propellant balance corresponding to each propulsion maneuver will be used as a future label.

[0073] During the inference phase, once the new driving action is completed and the corresponding feature extraction is finished, the latest... The combined feature vectors corresponding to each propulsion action form a new time-series sample sequence, which is then input into the trained deep learning prediction model to output the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period.

[0074] Please see Figure 3 , Figure 3A schematic diagram of the intelligent prediction system for propellant balance in a Hall thruster provided by this invention. This application also provides an intelligent prediction system for propellant balance in a Hall thruster, comprising: The data acquisition module is used to acquire the operating data of the target Hall thruster. The operating data includes at least tank pressure, tank temperature, valve status, discharge voltage, discharge current, thrust command, and operating mode. The graph construction module is used to extract the shutdown recovery segment for each propulsion action and construct a recovery dynamics graph based on each shutdown recovery segment; The feature extraction module is used to extract residual sensitive features based on the recovery kinetic maps and the benchmark recovery kinetic maps under the corresponding working conditions. The prediction module is used to assemble the remaining sensitive features of each propulsion action into a time-series sample sequence in chronological order, and input it into the deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0076] 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 smart prediction method for propellant balance in Hall thrusters, characterized in that, include: S1. Obtain the operating data of the target Hall thruster, wherein the operating data includes at least tank pressure, tank temperature, valve status, discharge voltage, discharge current, thrust command and operating mode; S2. Extract the shutdown recovery segment for each propulsion action, and construct a recovery dynamics map based on each shutdown recovery segment; S3. Based on the recovery dynamics maps described above and the baseline recovery dynamics maps under the corresponding working conditions, extract the residual sensitive feature quantities; S4. The remaining sensitive features of each propulsion action are arranged into a time series sample sequence in chronological order and input into the deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period.

2. The intelligent prediction method for propellant balance in Hall thrusters according to claim 1, characterized in that, The extraction shutdown recovery segment includes: Based on the valve status changing from open to closed and the discharge current dropping below the preset cutoff threshold, determine the shut-off starting point corresponding to each propulsion action; Data on the changes in tank pressure and tank temperature over time are extracted from each of the aforementioned shutdown points; If the rate of change of tank pressure is less than a first threshold and the rate of change of tank temperature is less than a second threshold within a continuous preset time period, then the start time of the continuous preset time period is taken as the end time of recovery. The data between each shutdown start point and the corresponding recovery end time is defined as the shutdown recovery segment.

3. The intelligent prediction method for propellant balance in Hall thrusters according to claim 2, characterized in that, The construction of the recovery dynamics map based on each of the shutdown recovery segments includes: The tank pressure and tank temperature data in each of the shutdown recovery sections are normalized to obtain normalized recovery data; Based on the normalized recovery data, a time-domain recovery subplot, a pressure-temperature phase diagram subplot, and a recovery rate subplot are constructed; wherein, the time-domain recovery subplot includes a normalized pressure recovery curve and a normalized temperature recovery curve, the pressure-temperature phase diagram subplot is a coupled trajectory formed by pressure and temperature in time sequence, and the recovery rate subplot includes a pressure recovery rate curve and a temperature recovery rate curve. The time-domain recovery sub-map, the pressure-temperature phase diagram sub-map, and the recovery rate sub-map are combined to obtain the recovery kinetics map.

4. The intelligent prediction method for propellant balance in Hall thrusters according to claim 3, characterized in that, The normalization process for the tank pressure and tank temperature data within each of the shutdown and recovery sections includes: Time normalization processing is performed on the time data corresponding to each sampling time within each of the aforementioned shutdown recovery segments; Based on the tank pressure corresponding to the shutdown start point and the tank pressure corresponding to the recovery end point in each shutdown recovery segment, the tank pressure data in each shutdown recovery segment are normalized to obtain a normalized pressure recovery curve. Based on the tank temperature corresponding to the shutdown start point and the tank temperature corresponding to the recovery end time in each shutdown recovery segment, the tank temperature data in each shutdown recovery segment are normalized to obtain a normalized temperature recovery curve. Based on the normalized pressure recovery curve and the normalized temperature recovery curve, the pressure recovery rate curve and the temperature recovery rate curve are obtained respectively.

5. The intelligent prediction method for propellant balance in Hall thrusters according to claim 1, characterized in that, The baseline recovery dynamics spectrum under the corresponding operating conditions includes: Historical propulsion actions are classified according to at least one of the following: thrust command level, operating mode, average discharge current range, average discharge voltage range, and propulsion duration range. Time alignment and statistical processing are performed on multiple historical recovery kinetic maps belonging to the same working condition category to obtain the baseline recovery kinetic map under the corresponding working condition; The reference recovery kinetic map includes at least a reference time domain recovery sub-map, a reference pressure-temperature phase map sub-map, and a reference recovery rate sub-map. The reference time domain recovery sub-map includes a reference pressure recovery curve and a reference temperature recovery curve, and the reference recovery rate sub-map includes a reference pressure recovery rate curve and a reference temperature recovery rate curve.

6. The intelligent prediction method for propellant balance in Hall thrusters according to claim 5, characterized in that, S3 includes: The deviation information is obtained by comparing the recovery dynamics map corresponding to the current propulsion action with the baseline recovery dynamics map under the corresponding working condition. The deviation information includes at least the difference information between the current time domain recovery subgraph and the reference time domain recovery subgraph, the geometric deviation information between the current pressure-temperature phase diagram subgraph and the reference pressure-temperature phase diagram subgraph, and the difference information between the current recovery rate subgraph and the reference recovery rate subgraph. Based on the recovery dynamics map corresponding to the current propulsion action and the deviation information, the residual sensitive feature quantity is extracted.

7. The intelligent prediction method for propellant balance in Hall thrusters according to claim 6, characterized in that, The margin-sensitive feature quantity includes at least one or more of the following: Pressure half-recovery time, temperature half-recovery time, hysteresis between pressure recovery peak and temperature recovery peak, area enclosed by pressure-temperature phase diagram, centroid coordinates of pressure-temperature phase diagram, principal axis direction angle of pressure-temperature phase diagram, integral difference between current normalized pressure recovery curve and reference pressure recovery curve, integral difference between current normalized temperature recovery curve and reference temperature recovery curve, difference between current pressure recovery rate curve and reference pressure recovery rate curve, and difference between current temperature recovery rate curve and reference temperature recovery rate curve.

8. The intelligent prediction method for propellant balance in Hall thrusters according to claim 1, characterized in that, S4 includes: The remaining sensitive features of each propulsion action and the propulsion condition parameters corresponding to each propulsion action are arranged in chronological order to form a time-series sample sequence. The time-series sample sequence is input into a deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period; The propulsion condition parameters include at least one or more of the following: propulsion duration, average discharge current, average discharge voltage, thrust command level, and operating mode. The deep learning prediction model is a long short-term memory network, a gated recurrent unit network, or a Transformer network.

9. An intelligent prediction system for propellant balance in Hall thrusters, characterized in that, include: The data acquisition module is used to acquire the operating data of the target Hall thruster. The operating data includes at least tank pressure, tank temperature, valve status, discharge voltage, discharge current, thrust command, and operating mode. The graph construction module is used to extract the shutdown recovery segment for each propulsion action and construct a recovery dynamics graph based on each shutdown recovery segment; The feature extraction module is used to extract residual sensitive features based on the recovery kinetic maps and the benchmark recovery kinetic maps under the corresponding working conditions. The prediction module is used to assemble the remaining sensitive features of each propulsion action into a time-series sample sequence in chronological order, and input it into the deep learning prediction model to obtain the current propellant remaining quantity prediction result and the propellant remaining quantity prediction result for the future period.

Images