Intelligent management system based on metal powder production

Abstract

The invention discloses an intelligent management system based on metal powder production, and relates to the technical field of industrial automation intelligent management. Sensor arrays are deployed in an ultrasonic atomizer and a vacuum drying furnace, the dust concentration, the vibration frequency, the pressure, the humidity, the temperature and the oxygen concentration are measured in real time, and the current stage is obtained through a production plan; dynamically and finely adjusting the weight matrix according to the sensor parameters corresponding to different stages, generating a weighted comprehensive index, calculating theoretical control parameters based on a physical model, correcting the theoretical control parameters through an LSTM neural network in combination with real-time environmental data, outputting an equipment control instruction, and extracting the sphericity and the moisture residue of the metal powder after the production of the metal powder is completed, so as to obtain the spherical degree of the metal powder. And the production quality is judged, and different responses are adopted according to the judgment result. The multi-parameter cooperative control precision is further improved, and the system stability is enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent management of industrial automation, and particularly relates to an intelligent management system based on metal powder production. BACKGROUND

[0002] Metal powder production is a core link in the fields of additive manufacturing and powder metallurgy, and the quality control precision of the metal powder production directly affects the performance of downstream products.

[0003] Traditional production systems usually adopt a fixed parameter control mode, and the differences in key stages such as atomization and drying are not optimized in stages, and the existing system lacks a stage-specific weight adjustment mechanism, resulting in insufficient multi-parameter collaborative control precision.

[0004] Based on the physical model constructed based on fluid mechanics and heat and mass transfer theory, it is easy to cause deviation between theoretical parameters and actual working conditions, and in the prior art, model parameter correction depends on artificial experience or offline calibration, and cannot capture nonlinear dynamics in real time, resulting in lag of control instructions and poor product quality stability.

[0005] Therefore, there is an urgent need for a management system to provide a new technical path for the intelligentization and flexibility of metal powder production. SUMMARY

[0006] In view of the defects of the prior art, the application provides an intelligent management system based on metal powder production, which solves the problems of insufficient multi-stage parameter collaboration, large model deviation and quality feedback lag.

[0007] To achieve the above purpose, the application is implemented by the following technical scheme: an intelligent management system based on metal powder production, comprising:

[0008] A multi-sensor end, a sensor array is arranged at an ultrasonic atomizer and a vacuum drying furnace, and is used for real-time dust concentration, vibration frequency, pressure, humidity, temperature and oxygen concentration.

[0009] An edge computing end acquires the current stage through a production plan, dynamically adjusts a weight matrix according to each sensor parameter corresponding to different stages, and generates a weighted comprehensive index.

[0010] A parameter correction end calculates theoretical control parameters based on a physical model, and corrects the theoretical control parameters by combining real-time environmental data through an LSTM neural network, and outputs device control instructions.

[0011] A quality feedback end extracts the sphericity and moisture residue of the metal powder after the production is completed, judges the production quality, and takes different responses according to the judgment result.

[0012] As a further scheme of the present application, the sensor array deployed in the ultrasonic atomizer specifically comprises a dust concentration sensor, a vibration frequency sensor, and a pressure sensor, the sensor array deployed in the vacuum drying furnace specifically comprises a temperature and humidity sensor, a heating element temperature sensor, and an oxygen sensor, and the sensor array deployed in the transmission pipeline specifically comprises a temperature and humidity sensor and a dust concentration sensor.

[0013] As a further scheme of the present application, the current stage specifically comprises an atomization stage and a drying stage.

[0014] As a further scheme of the present application, the initial weight matrix corresponding to the atomization stage is Watomization=[0.5, 0.3, 0.2], corresponding to the dust concentration, the vibration frequency, and the pressure respectively, and the initial weight matrix corresponding to the drying stage is Wdrying=[0.6, 0.3, 0.1], corresponding to the humidity, the temperature, and the oxygen concentration respectively.

[0015] As a further scheme of the present application, the specific steps of dynamically fine-tuning the weight matrix are as follows:

[0016] A threshold value xth of each sensor parameter x is preset, and the allowed fluctuation range is [xth-D, xth+D], wherein D is a threshold value buffer value;

[0017] Real-time data xreal of each sensor is obtained, and an absolute difference value d=|xreal-xth| with the threshold value is calculated;

[0018] The initial weight matrix is W=[w1, w2, wx];

[0019] If d<=D, the weight fine-tuning is started;

[0020] The fine-tuned weight is calculated according to the formula Δw=k·(1-(d / D)^2), wherein k is an adjustment coefficient;

[0021] The fine-tuned sensor weight is calculated according to the formula wx'=wx+Δx;

[0022] The overall fine-tuned weight matrix is obtained according to the formula

[0023] The initial weight matrix specifically refers to the initial weight matrix corresponding to the atomization and drying stages, and the real-time dust concentration, vibration frequency, and pressure are collected in the atomization stage, and the real-time humidity, temperature, and oxygen concentration are collected in the drying stage.

[0024] As a further scheme of the present application, the weighted comprehensive index is calculated according to the formula S=∑(W norm ×Normalize(d i )), wherein di is the original sensor data, and Normalize() is a normalization function.​

[0025] As a further scheme of the present application, the specific steps for calculating the theoretical control parameters based on the physical model are:

[0026] The theoretical optimal vibration frequency is calculated according to the formula wherein k is a constant, d is the diameter of the nozzle, σ is the surface tension, and ρ is the liquid density;

[0027] The theoretical temperature is calculated according to the formula wherein T0 is the initial temperature, Q is the heating power, t is the time, m is the mass of the material, and cp is the specific heat capacity.

[0028] As a further scheme of the present application, the specific steps for correcting the theoretical values using LSTM are:

[0029] Historical data of the atomization stage is collected, including vibration frequency, pressure, dust concentration, and the theoretical optimal vibration frequency ftheoretical and frequency residual based on the physical model;

[0030] Historical data of the drying stage is collected, including humidity, temperature, oxygen concentration, and the theoretical temperature Ttheoretical and temperature residual based on the physical model;

[0031] The data of the two stages are respectively standardized and processed for outliers, and training samples are generated according to the time window;

[0032] In the atomization stage, the input parameters of LSTM are vibration frequency, pressure, dust concentration, theoretical optimal vibration frequency ftheoretical, and weighted comprehensive index Satomization, and the output is frequency residual;

[0033] The specific parameters of the LSTM parameters include 1 layer of LSTM, [32, 64] number of neurons, selection of Adam optimizer, learning rate of 0.001, batch size of [16, 32], and training rounds of [50, 100];

[0034] In the drying stage, the input parameters of LSTM are humidity, temperature, oxygen concentration, theoretical temperature Ttheoretical, and weighted comprehensive index Sdry, and the output is temperature residual;

[0035] The specific parameters of the LSTM parameters include 2 layers of LSTM, [64, 128] number of neurons, selection of AdamW optimizer, learning rate of 0.0005, batch size of [32, 64], and training rounds of [100, 200];

[0036] After the training is completed, an atomization LSTM model and a drying LSTM model are obtained, different LSTM models are selected according to the stage, and frequency residual and temperature residual are obtained respectively;

[0037] The frequency residual, the temperature residual and the corresponding theoretical optimal vibration frequency ftheoretical and the theoretical temperature Ttheoretical are added to obtain the corrected optimal vibration frequency and temperature.

[0038] As a further scheme of the present application, the specific steps of judging the production quality are:

[0039] If the detection result is high sphericity and extremely low residual moisture, it is judged as qualified; if it is medium sphericity or trace residual, it is judged as single-stage slight deviation; and if it is irregular sphericity or obvious residual, it is judged as significant deviation.

[0040] As a further scheme of the present application, the specific steps of taking different response strategies according to the judgment result are:

[0041] If it is judged as qualified, the current stage parameter combination is solidified and used as the default production strategy for the production of the same type of metal powder;

[0042] If it is judged as single-stage slight deviation and the particle sphericity is medium sphericity, the historical medium sphericity quality adjustment scheme is recalled, and the atomization LSTM model is incrementally trained;

[0043] If it is judged as single-stage slight deviation and the particle sphericity is trace residual, the hygroscopicity parameter of the drying mechanism model is increased;

[0044] If it is judged as significant deviation, the man-machine collaborative mechanism is started, the material pretreatment link is manually rechecked, the LSTM model state of the corresponding stage is reset synchronously, and the interference of abnormal data on the model is removed.

[0045] The present application provides an intelligent management system for metal powder production, which has the following advantages compared with the prior art:

[0046] (1) The present application dynamically fine-tunes the sensor weight matrix by the edge computing end in stages, generates a weighted comprehensive index, and makes the dust concentration, vibration frequency and pressure in the atomization stage and the humidity, temperature and oxygen concentration in the drying stage accurately coupled according to the stage characteristics, improves the multi-parameter collaborative control precision, and enhances the system stability;

[0047] (2) The present application calculates the theoretical parameters based on the physical model, combines the LSTM neural network to correct the deviation between the theoretical value and the actual working condition in real time, makes up for the defects of the simplified assumptions of the traditional model, reduces the quality fluctuation caused by the model deviation, and improves the accuracy of the control command;

[0048] (3) The application classifies and responds by detecting the sphericity and moisture residue in real time, dynamically correlates the quality detection result with the parameters in the atomization and drying stages, realizes rapid positioning of abnormal links, and reduces the cost of manual intervention. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 The figure is a system principle block diagram of the application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0051] As Figure 1 The application provides an intelligent management system based on metal powder production, which comprises:

[0052] A multi-sensor end, wherein a dust concentration sensor and a vibration frequency sensor are arranged near the spray head of the ultrasonic atomizer, and a pressure sensor is arranged inside the cavity.

[0053] During the ultrasonic atomization process, suspended dust is generated when the metal molten liquid is broken into micron-sized particles, and there is a safety hazard when the concentration exceeds the lower explosion limit. At the same time, too high dust concentration can cause the spray head to be blocked, affecting the consistency of particle size.

[0054] The real-time dust concentration is collected by the dust concentration sensor.

[0055] The core parameter of ultrasonic atomization is vibration frequency. According to the formula d approximately equal to k times c divided by f, it can be seen that the size of the vibration frequency directly determines the particle size of the liquid droplet breaking. In the above formula, k is a constant, c is the speed of sound, and f is the vibration frequency.

[0056] The frequency may drift due to crystal aging and mechanical fatigue after long-term work of the spray head, and needs to be calibrated in real time.

[0057] The real-time vibration frequency is collected by the vibration frequency sensor.

[0058] The atomization cavity needs to maintain stable negative pressure to prevent metal vapor leakage, and at the same time, the negative pressure environment can reduce the oxidation rate of particles. Fluctuation of the gas supply pressure will cause unstable atomization liquid supply, forming satellite particles, i.e. small particles around the main particles.

[0059] The absolute pressure of the cavity is collected by the pressure sensor.

[0060] Deploy temperature and humidity sensors, heating element temperature sensors, and oxygen sensors inside the vacuum drying furnace to form a distributed monitoring network;

[0061] Humidity is the main factor causing metal powder oxidation, while temperature directly affects drying efficiency and oxidation rate. High temperature may accelerate oxidation;

[0062] Real-time collection of relative humidity in the furnace by temperature and humidity sensors;

[0063] The actual temperature of the heating element may deviate from the set temperature, such as due to aging, which may reduce the heating efficiency and directly affect the uniformity of powder heating;

[0064] Collecting the surface temperature of the heating element by the heating element temperature sensor to verify the stability of the heating system;

[0065] In a vacuum environment, trace amounts of oxygen may still cause metal powder oxidation, especially at high temperatures where oxidation reactions intensify;

[0066] Collecting the oxygen concentration in the furnace by the oxygen sensor to monitor whether the vacuum degree meets the standard and whether there is a leak;

[0067] Deploy temperature and humidity and dust concentration sensors every 5-10 meters in the transmission pipeline to achieve full-process coverage of the production environment;

[0068] Metal powder may change its physical properties during transmission due to changes in environmental parameters such as temperature and humidity, dust concentration, and the specific effects are as follows:

[0069] Temperature and humidity: increased humidity may cause strong hygroscopic metal powder to clump, agglomerate, block the pipeline, or affect the accuracy of subsequent screening;

[0070] Temperature fluctuations may cause powder thermal expansion and contraction, changing the bulk density and causing unstable transmission flow, or even causing local oxidation due to frictional heating;

[0071] Dust concentration: high concentration of suspended dust may form an explosive environment, especially for active metal powders such as magnesium and aluminum, which need to be monitored in real-time to prevent the risk of explosion;

[0072] Dust deposition may cause increased wear and tear of the inner wall of the pipeline, affecting transmission efficiency.

[0073] Edge computing, through production planning, automatically adjusts the weight of each sensor parameter according to the stage to generate a comprehensive environmental state index;

[0074] The stage specifically includes two key stages: atomization and drying;

[0075] The reasons for choosing the atomization stage are as follows:

[0076] The atomization stage is the core stage that determines the physical properties of metal powder particles. In the ultrasonic atomization process, parameters such as vibration frequency and pressure directly affect the droplet breaking effect, and thus determine the particle size distribution and morphology characteristics of the powder. These characteristics are key quality indicators for downstream applications.

[0077] The equipment needs to run continuously at high power during this stage, and energy consumption is closely related to production load. Traditional control methods can easily lead to energy efficiency reduction when the load changes.

[0078] Natural degradation of equipment performance (such as changes in transducer efficiency) or fluctuations in raw material properties (such as differences in the physical properties of molten liquid) can cause parameter drift, which in turn affects atomization stability. By monitoring and dynamically calibrating parameters in real time, particle size uniformity can be ensured, and subsequent stage efficiency decline or product defects can be avoided.

[0079] The reasons for choosing the drying stage are as follows:

[0080] The drying stage is a critical link for moisture removal and oxidation protection in metal powder production. Metal powder is extremely sensitive to changes in humidity and oxygen concentration in a high-temperature environment. Residual moisture can affect the stability of subsequent processing, and the presence of oxygen can cause surface oxidation of the powder, which in turn affects material performance.

[0081] The heating and vacuum equipment involved in this stage consumes a large proportion of energy. Traditional fixed parameter control methods cannot adapt to environmental fluctuations, which can lead to energy waste or incomplete drying.

[0082] There is a complex coupling relationship between parameters such as temperature, humidity, and vacuum degree. Small changes in environmental conditions can trigger a chain reaction, such as changes in temperature affecting the sealing of the equipment, leading to the infiltration of oxygen. Dynamic data fusion and real-time control models are needed to achieve precise regulation and ensure powder quality from the source.

[0083] The weight matrix of the pre-set drying stage and atomization stage is as follows:

[0084] For the atomization stage, the weight matrix is set as Watomization = [0.5, 0.3, 0.2], corresponding to dust concentration, vibration frequency, and pressure, respectively.

[0085] Dust concentration is a comprehensive reflection of particle generation efficiency and uniformity during atomization. It essentially reflects the number of effective broken droplets per unit volume and the concentration of particle size distribution. If the concentration is abnormal, it may indicate that the droplet breaking is insufficient or excessive, which will cause the particle size distribution to widen and affect the forming precision of processes such as additive manufacturing. As a result, it has the highest weight in terms of direct impact on quality.

[0086] Vibration frequency is the core driving parameter of ultrasonic atomization, directly determines the initial energy input of droplet breakage, the frequency is negatively correlated with particle size, and the stability directly affects the consistency of particle size distribution. Frequency micro-fluctuation can be dynamically corrected through dust concentration feedback, and the frequency stability of the equipment itself is usually high, so the weight is lower than that of dust concentration;

[0087] Pressure is an auxiliary energy parameter of atomization process, mainly to maintain the stability of liquid flow. When the vibration frequency is stable, the pressure fluctuation has a nonlinear effect on particle characteristics, but its direct regulation on particle size is weaker than that of vibration frequency. It indirectly affects the atomization effect through the coupling relationship with frequency, so the weight is the lowest;

[0088] For the drying stage, the weight matrix is specifically set as Wdry = [0.6, 0.3, 0.1], corresponding to air relative humidity, heating element surface temperature, and oxygen concentration, respectively;

[0089] Humidity is the most direct parameter to determine the residual moisture content of powder. Moisture residue can cause defects such as pores and cracks in the subsequent processing process, directly affecting material performance. At the same time, the risk of powder oxidation increases exponentially with every threshold increase in humidity, and the effect of humidity fluctuation on drying efficiency has a lagging nature, which needs to be dynamically responded to first;

[0090] Temperature directly controls the drying rate and uniformity. Insufficient temperature will lead to incomplete drying, and excessive temperature may exacerbate powder surface oxidation. However, temperature regulation has instantaneity, and its influence can be indirectly corrected through humidity feedback, so the weight is lower than that of humidity;

[0091] In the drying stage, oxygen concentration is usually controlled at a low level through sealing design, and its effect on oxidation is more potential risk. In contrast, the real-time change of oxygen concentration has a weaker direct impact on the quality of the current stage than humidity and temperature, so the weight is the lowest;

[0092] After determining the stage, select the weight matrix corresponding to each stage, and dynamically fine-tune the weight matrix according to the sensor data fluctuation;

[0093] When the sensor data approaches the threshold, the system needs to give an early warning and enhance the response sensitivity. Through weight enhancement, the influence of this variable in the control algorithm can be amplified;

[0094] The specific steps of dynamically fine-tuning the weight matrix according to the sensor data fluctuation are as follows:

[0095] The threshold of a certain sensor parameter x is xth, and the allowed fluctuation range is [xth-D, xth+D], where D is the threshold buffer value;

[0096] The sensor real-time data is obtained as xreal, and the absolute difference d = |xreal-xth| is calculated;

[0097] If the initial weight matrix is W = [w1, w2, wx];

[0098] If d <= D, it indicates that the parameter is close to the threshold, and the weight is promoted;

[0099] The fine-tuned weight is calculated according to the formula Δw = k·(1-(d / D)^2), where k is the adjustment coefficient;

[0100] The fine-tuned sensor weight is calculated according to the formula wx' = wx+Δx;

[0101] The overall fine-tuned weight matrix is obtained according to the formula

[0102] For example, the weight matrix of the drying stage is Wdry = [0.6, 0.3, 0.1], corresponding to humidity, temperature, and oxygen concentration, respectively. The relative humidity of the air in the oven is collected by the temperature and humidity sensor in real time as 25% RH, the humidity threshold xth is 30% RH, and the threshold buffer value D is 10%. The absolute difference d = |25%-30%| = 5% <10% is calculated, triggering the weight fine-tuning, while the temperature and oxygen concentration do not meet the triggering condition. At this time, the fine-tuned weight is calculated according to the formula Δw = k·(1-(d / D)^2), where k is set to 0.1, then Δw = 0.075 is calculated, and the fine-tuned weight matrix is [0.675, 0.3, 0.1], the total is 1.075, which exceeds 1, and needs to be normalized to obtain the fine-tuned weight matrix [0.628, 0.279, 0.093];

[0103] After obtaining the fine-tuned weight matrix of different stages, the weighted comprehensive index is calculated according to the formula S = ∑(W norm ×Normalize(d i )), where di is the original sensor data, and Normalize() is the normalization function, which unifies the sensor data to [0, 1];

[0104] By converting multi-dimensional sensor data into a single quantifiable comprehensive index, subsequent system processing and decision-making are facilitated.

[0105] The parameter correction end calculates the theoretical control parameter based on the physical model calculation theory, and corrects the parameter through the LSTM neural network combined with real-time environmental data, and outputs the device control instruction;

[0106] Based on the fluid mechanics equation, the theoretical best vibration frequency is calculated according to the formula , where k is a constant, d is the diameter of the nozzle, σ is the surface tension, and ρ is the liquid density;​

[0107] Based on Fourier heat equation and Fick diffusion law, the theoretical temperature Ttheoretical is calculated according to the formula wherein T0 is the initial temperature, Q is the heating power, t is the time, m is the mass of the material, and cp is the specific heat capacity;

[0108] Since the physical model is based on simplified assumptions, such as ignoring nonlinear characteristics and idealizing boundary conditions, the values obtained have deviations from the actual complex working conditions. At the same time, the parameters of the physical model may drift over time, so it is necessary to use LSTM to correct the theoretical values;

[0109] The specific steps of using LSTM to correct the theoretical values are as follows:

[0110] Collect historical data of the atomization tower, including vibration frequency, pressure, dust concentration, and the theoretical optimal vibration frequency ftheoretical and frequency residual based on the physical model;

[0111] Collect historical data of the drying furnace, including humidity, temperature, oxygen concentration, and the theoretical temperature Ttheoretical and temperature residual based on the physical model;

[0112] Standardize and process outliers for the two-stage data, and generate training samples according to the time window, such as 1 minute / window for the atomization stage and 10 minutes / window for the drying stage;

[0113] In the atomization stage, the input parameters of LSTM are vibration frequency, pressure, dust concentration, theoretical optimal vibration frequency ftheoretical, and weighted comprehensive index Satomization, and the output is frequency residual;

[0114] The specific values of the LSTM parameters are as follows:

[0115] The number of layers of LSTM is 1, and the physical mechanism of the atomization stage is relatively clear, so a single layer can capture the main time sequence association;

[0116] The number of neurons is [32, 64], and the vibration frequency is affected by limited factors, so a medium-capacity model is sufficient to fit the nonlinear relationship and avoid overfitting;

[0117] Select the Adam optimizer and set the learning rate to 0.001 to balance the convergence speed and stability;

[0118] The batch size is [16, 32], and the data in the atomization stage is real-time, so small batches are suitable for online training or edge computing;

[0119] The number of training rounds is [50, 100], combined with the early stopping mechanism, to avoid overtraining the rules covered by the physical model;

[0120] In the drying stage, the input parameters of LSTM are humidity, temperature, oxygen concentration, theoretical temperature Ttheoretical, and weighted comprehensive index Sdrying, and the output is temperature residual;

[0121] The specific values of LSTM parameters are as follows:

[0122] The number of layers of LSTM is 2, the multivariate coupling in the drying stage is strong, and the deep network can capture long-term dependence and complex nonlinear relationship;

[0123] The number of neurons is [64, 128], the first layer is used to capture basic features, and the second layer is used to integrate high-order correlation;

[0124] AdamW is selected with weight decay to prevent overfitting of deep network, and the learning rate is 0.0005, which is lower to stabilize the training of deep network;

[0125] The batch size is [32, 64], the data in the drying stage may be more abundant, and larger batch size improves training stability;

[0126] The number of training rounds is [100, 200], combined with the early stopping mechanism, allowing the deep network to fully learn the multivariate coupling rule;

[0127] After training, the atomization LSTM model and the drying LSTM model are obtained, and different LSTM models are selected according to the stage to obtain frequency residual and temperature residual, respectively, and the corresponding theoretical best vibration frequency ftheoretical and theoretical temperature Ttheoretical are added to obtain the modified best vibration frequency and temperature.

[0128] Quality feedback end, after the production of metal powder, extract its sphericity and moisture residue, and judge its production quality, according to the judgment result, take phased response;

[0129] Particle sphericity directly affects the flowability and forming precision of metal powder in subsequent applications such as additive manufacturing;

[0130] Moisture residue grade is related to the stability and chemical activity of powder storage, which needs to ensure that it meets the quality standards;

[0131] Both indicators are key to measuring the quality of metal powder, in addition to these two indicators, which correspond to the atomization and drying stages respectively, can quickly locate the abnormal production link, and facilitate the following process parameter optimization;

[0132] Detect particle sphericity through image recognition algorithm, mark it as high sphericity (good), medium sphericity (medium), and irregular sphericity (poor);

[0133] The industrial camera is used to dynamically shoot the laid metal powder particles. After the particle images are acquired, a machine learning algorithm (such as a convolutional neural network CNN) or a traditional image processing algorithm (such as edge detection and contour extraction) is used to calculate the shape parameters of the particles, such as the circularity coefficient, the aspect ratio and the elongation rate, and compare them with the parameter threshold of the standard sphere to quantify the sphericity level.

[0134] The detection method can be integrated at the end of the production line to realize online dynamic detection without manual sampling or offline analysis. It can directly interface with the production control system and instantly feedback the sphericity abnormality. Since there is no physical contact with the particles, the detection process does not damage the surface state of the powder, and it is especially suitable for easily oxidized metal powder.

[0135] The residual moisture is detected by near-infrared spectroscopy and marked as extremely low residual (excellent), trace residual (medium) and obvious residual (poor).

[0136] The metal powder sample is irradiated with near-infrared light. The O-H bond in the water molecule will absorb light of a specific wavelength. The transmittance or reflectance spectrum is collected by a spectrometer, and a chemometrics model such as partial least squares (PLS) is used to establish a quantitative relationship between the spectral characteristics and the moisture content, and directly output the moisture residual level.

[0137] The detection method takes only a few seconds to tens of seconds per detection, and does not damage the sample, making it suitable for real-time monitoring in continuous production processes, such as installing an online spectral probe at the outlet of the drying stage to update moisture data every 10 seconds.

[0138] If the detection result is high sphericity and extremely low moisture residual, it is judged to be qualified. If it is medium sphericity or trace residual, it is judged to be a slight deviation in a single stage. If it is irregular sphericity or obvious residual, it is judged to be a significant deviation.

[0139] The specific steps for taking different response strategies according to the judgment result are as follows:

[0140] If it is judged to be qualified, the current stage parameter combination is fixed and used as the default production strategy for the production of similar metal powder.

[0141] If it is judged to be a slight deviation in a single stage and the particle sphericity is medium sphericity (i.e., there is a problem in the atomization stage), the historical medium sphericity quality adjustment scheme is retrieved, and the LSTM model for atomization is incrementally trained to find the best vibration frequency correction value.

[0142] If it is judged to be a slight deviation in a single stage and the particle sphericity is trace residual (i.e., there is a problem in the drying stage), the moisture absorption rate parameter of the drying mechanism model is increased. For example, the theoretical moisture absorption rate is adjusted from 0.8 to 0.85 according to the current material characteristics.

[0143] If it is judged that there is a significant deviation, a man-machine cooperation mechanism is started, the material pretreatment link is manually rechecked, whether the material is mixed uniformly before drying is checked, the LSTM model state of the corresponding stage is reset, and the interference of abnormal data on the model is cleared.

[0144] Some data in the above formula are dimensionless for numerical calculation, and the contents not described in detail in the specification all belong to the prior art known to those skilled in the art.

[0145] The above examples are only used to illustrate the technical method of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. An intelligent management system for metal powder production, characterized in that: include: Multi-sensor end: sensor arrays are deployed in ultrasonic atomizers and vacuum drying ovens to monitor real-time dust concentration, vibration frequency, pressure, humidity, temperature, and oxygen concentration. The edge computing end obtains the current stage through the production plan, dynamically adjusts the weight matrix according to the sensor parameters corresponding to different stages, and generates a weighted comprehensive index; The parameter correction end calculates theoretical control parameters based on the physical model, and corrects the theoretical control parameters through the LSTM neural network combined with real-time environmental data, and outputs device control instructions; At the quality feedback end, after the metal powder production is completed, its sphericity and moisture residue are extracted, and the production quality is judged, and different responses are taken according to the judgment results.

2. The intelligent management system for metal powder production according to claim 1, characterized in that: The sensor array deployed in the ultrasonic atomizer specifically includes a dust concentration sensor, a vibration frequency sensor, and a pressure sensor; the sensor array deployed in the vacuum drying furnace specifically includes a temperature and humidity sensor, a heating element temperature sensor, and an oxygen sensor; the sensor array deployed in the transmission pipeline specifically includes a temperature and humidity and dust concentration sensor.

3. The intelligent management system for metal powder production according to claim 1, characterized in that: The current stage specifically includes an atomization stage and a drying stage.

4. The intelligent management system for metal powder production according to claim 3, characterized in that: The initial weight matrix corresponding to the atomization stage is Watomization = [0.5, 0.3, 0.2], which corresponds to dust concentration, vibration frequency, and pressure, respectively; the initial weight matrix corresponding to the drying stage is Wdrying = [0.6, 0.3, 0.1], which corresponds to humidity, temperature, and oxygen concentration, respectively.

5. The intelligent management system for metal powder production according to claim 1, characterized in that: The specific steps for dynamically fine-tuning the weight matrix are: The threshold xth of each sensor parameter x is preset, and the allowed fluctuation range is [xth-D,xth+D], where D is the threshold buffer value; Get the real-time data xreal of each sensor and calculate the absolute difference with the threshold d = |xreal-xth|; The initial weight matrix is ​​W = [w1, w2, wx]; If d <= D, start weight fine-tuning; The fine-tuning weight is calculated using the formula Δw = k (1-(d / D)^2), where k is the adjustment coefficient. Calculate the sensor weight after fine-tuning according to the formula wx'=wx+Δx; According to the formula Get the weight matrix after overall fine-tuning; The initial weight matrix is ​​specifically an initial weight matrix corresponding to the atomization and drying stages. The atomization stage collects real-time dust concentration, vibration frequency, and pressure, and the drying stage collects real-time humidity, temperature, and oxygen concentration.

6. The intelligent management system for metal powder production according to claim 1, characterized in that: According to the formula S=∑(W norm ×Normalize(d i ))Calculate the weighted comprehensive index, where di is the original sensor data and Normalize() is the normalization function.

7. The intelligent management system for metal powder production according to claim 1, characterized in that: The specific steps for calculating theoretical control parameters based on the physical model are: According to the formula Calculate the theoretical optimal vibration frequency, where k is a constant, d is the nozzle diameter, σ is the surface tension, and ρ is the liquid density; According to the formula Calculate the theoretical temperature, where T0 is the initial temperature, Q is the heating power, t is the time, m is the material mass, and cp is the specific heat capacity.

8. The intelligent management system for metal powder production according to claim 1, characterized in that: The specific steps for correcting the theoretical value using LSTM are: Collect historical data of the atomization stage, including vibration frequency, pressure, dust concentration, and the theoretical optimal vibration frequency ftheoretical and frequency residual obtained based on the physical model; Collect historical data of the drying stage, including humidity, temperature, oxygen concentration, as well as theoretical temperature Ttheoretical and temperature residual obtained based on the physical model; The two-stage data are standardized and outlier processed respectively, and training samples are generated according to the time window; In the atomization stage, the input parameters of LSTM are vibration frequency, pressure, dust concentration, theoretical optimal vibration frequency ftheoretical, and weighted comprehensive index Satomization, and the output is frequency residual; The specific parameters of LSTM parameters include the number of LSTM layers as 1, the number of neurons as [32, 64], the selection of Adam optimizer, the learning rate as 0.001, the batch size as [16, 32], and the number of training rounds as [50, 100]; In the drying stage, the input parameters of LSTM are humidity, temperature, oxygen concentration, theoretical temperature Ttheoretical, and weighted comprehensive index Sdrying, and the output is the temperature residual; The specific parameters of LSTM parameters include the number of LSTM layers as 2, the number of neurons as [64, 128], the selection of AdamW optimizer, the learning rate as 0.0005, the batch size as [32, 64], and the number of training rounds as [100, 200]; After training is completed, the atomized LSTM model and the drying LSTM model are obtained. Different LSTM models are selected according to the stage to obtain the frequency residual and temperature residual respectively; The frequency residual, temperature residual, corresponding theoretical optimal vibration frequency ftheoretical, and theoretical temperature Ttheoretical are added to obtain the corrected optimal vibration frequency and temperature.

9. The intelligent management system for metal powder production according to claim 1, characterized in that: The specific steps to judge production quality are: If the test result shows high sphericity and extremely low residual moisture, it is judged to be qualified; if it is medium sphericity or trace residual, it is judged to be a single-stage slight deviation; if it is irregular sphericity or obvious residual, it is judged to be a significant deviation.

10. The intelligent management system for metal powder production according to claim 1, characterized in that: The specific steps for adopting different response strategies based on the judgment results are as follows: If it is judged to be qualified, the current stage parameter combination is solidified and used as the default production strategy for the production of similar metal powders; If it is judged to be a slight deviation in a single stage and the particle sphericity is medium spherical, the historical medium spherical quality adjustment plan is called up, and at the same time, the atomization LSTM model is incrementally trained; If it is judged to be a slight deviation in a single stage and the particle sphericity is a trace residual, the moisture absorption rate parameter of the drying mechanism model is increased; If a significant deviation is determined, the human-machine collaboration mechanism will be activated to manually re-inspect the material pre-processing process, and the LSTM model status of the corresponding stage will be reset synchronously to eliminate the interference of abnormal data on the model.

Images