HPDI dual-fuel injection control system and method based on rail pressure signal and neural network

Abstract

The invention discloses an HPDI dual-fuel injection control system and method based on a rail pressure signal and a neural network, and relates to the technical field of engine fuel injection control. The collected rail pressure signals are preprocessed, conversion processing of converting time domain signals into frequency domain signals is executed to obtain frequency domain characteristics, and frequency band energy in a specific frequency range is extracted; according to the HPDI dual-fuel injection control system and method based on the rail pressure signal and the neural network, the new oil injector and the abraded oil injector correspond to different initial coefficients respectively, and can well adapt to different abrasion stages of the oil injectors; meanwhile, in order to effectively solve the problem that a traditional method is poor in adaptability to complex working conditions and the dynamic state of the fuel injector, the collected rail pressure signals cover multiple working conditions such as engine idling and acceleration, the neural network model can fully learn different working condition characteristics, and the system can stably operate in complex scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engine fuel injection control, and particularly relates to an HPDI dual-fuel injection control system and method based on a rail pressure signal and a neural network. BACKGROUND

[0002] In the scenarios of heavy trucks shuttling in intercity high-speed, engineering machinery continuously operating in the construction site, and ships sailing in the vast sea, the engine as the core power source, the precision of its fuel injection directly determines the stability of power output, the compliance of tail gas emission, and the economic utilization efficiency of fuel. The HPDI dual-fuel injection technology has become the mainstream advanced fuel supply method at present due to its high efficiency and energy saving advantage. Especially in the dual-fuel combination application of diesel and natural gas, the injection quantities of the two fuels need to be accurately controlled to flexibly adapt to different operating conditions of the engine from idle start, sudden acceleration overtaking to high-speed uniform driving, and to cope with the inevitable wear and tear of the injector in long-term high-frequency work.

[0003] However, the traditional injection control method has many limitations in actual application. For example, when the truck is climbing, the engine load increases suddenly, and the traditional control algorithm often causes the injection quantity to be not adjusted in time due to response lag, resulting in power jam. When the engineering machinery is suddenly overloaded after a long time of idle standby, the actual injection quantity deviates from the theoretical value due to slight wear of the injector, and the traditional method is difficult to quickly correct, which not only increases fuel consumption, but also may cause emission to exceed the standard. The root cause of these problems lies in the fact that the traditional method is difficult to achieve a good balance between control precision and adaptability to complex working conditions and dynamic state of the injector.

[0004] The particularity of the dual-fuel system aggravates the precision crisis: in the diesel / natural gas dual-fuel high-pressure direct injection engine, diesel only bears the micro-pilot action (typical injection quantity 2-10 mm 3 / stroke), and natural gas is the main fuel. There are double precision difficulties under this special working condition:

[0005] Micro-injection misalignment: when the diesel injection quantity is extremely small, the non-linear characteristics of the traditional injector cause the deviation to be amplified (more than ±20%);

[0006] Fuel ratio out of control: the actual injection ratio of dual fuel deviates from the instruction value, causing combustion instability:

[0007] (1) Excessive natural gas → unburned hydrocarbons (uHC) surge;

[0008] (2) Excessive diesel → local knock tendency rises;

[0009] Extreme working conditions worsen:

[0010] (1) Cold start stage: trace diesel injection deviation causes misfire;

[0011] (2) High load condition: fuel ratio imbalance leads to in-cylinder pressure oscillation;

[0012] Therefore, an intelligent closed-loop correction mechanism needs to be developed, which requires: adaptive compensation for injector wear throughout its life cycle to ensure dual-fuel ratio control accuracy within ±3%, meeting the national VIb and above emission regulations. For this purpose, we propose a HPDI dual-fuel injection control system and method based on rail pressure signals and neural networks. SUMMARY

[0013] (I) Technical problems solved

[0014] To address the shortcomings of existing technologies, the present application provides a HPDI dual-fuel injection control system and method based on rail pressure signals and neural networks, which solves the technical problem of traditional injection control methods being difficult to balance control accuracy and adaptability.

[0015] (ii) Technical solutions

[0016] To achieve the above objectives, the present application is implemented by the following technical solutions:

[0017] The HPDI dual-fuel injection control system based on rail pressure signals and neural networks comprises:

[0018] A rail pressure signal acquisition component for acquiring real-time rail pressure signals of the engine high-pressure common rail system;

[0019] A signal processing and analysis component connected to the rail pressure signal acquisition component, which preprocesses the collected rail pressure signals, performs transformation processing to convert time-domain signals to frequency-domain signals to obtain frequency-domain features, and extracts the frequency band energy E within a specific frequency range;

[0020] A neural network calculation component connected to the signal processing and analysis component, which receives the processed frequency band energy at its input part, processes it through a pre-set neural network structure, and outputs the model injection amount from the output part. The calculation method of the output part includes coefficients that can be updated online;

[0021] A closed-loop adjustment component connected to the neural network calculation component and the engine injector, respectively, receives the command injection amount and the model injection amount, calculates the deviation between the two, and dynamically adjusts the power-on time of the injector through the control device according to the deviation.

[0022] Preferably, the rail pressure signal acquisition component uses a high-precision pressure sensor that acquires 10,000 data points per second, and the acquisition range covers the rail pressure interval of 0 to 200 MPa;

[0023] The preprocessing of the signal processing and analysis component includes removing impulse noise and high-frequency interference in the signal, and performing smoothing filtering. The conversion of the time-domain signal to the frequency-domain signal is performed on every 1024 continuous data points, which correspond to a time window of 102.4 ms, and the specific frequency range extracted is 100 Hz to 500 Hz.

[0024] The preprocessing of the rail pressure signal by the signal processing and analysis component also includes temperature compensation of the signal, which corrects the collected signal according to the working temperature of the sensor.

[0025] Preferably, the preset neural network structure of the neural network calculation component includes two fully connected hidden parts, each of which includes 16 processing units, adopts a linear rectification operation as the activation mode, and the weight matrix is 1 row by 16 columns and 16 rows by 16 columns, respectively, and the initial value of the weight matrix is obtained through offline training; the calculation formula of the output part is:

[0026] Q model =k×ln(E+c)+b;

[0027] In the formula, Q model is the model injection amount, E is the frequency band energy, k, c, and b are three constant coefficients.

[0028] Preferably, the three coefficients in the neural network calculation component that can be updated online are updated through online learning to adapt to different wear stages of the fuel injector. In the initial coefficients of the new fuel injector, k is 0.18, b is 12.7, and c is 5.2.

[0029] In the coefficients of the worn fuel injector, k is 0.22, b is 14.1, and c is 6.8.

[0030] During online learning, the coefficients are calibrated and updated once every 100 hours of operation. During the updating process, the coefficients are corrected by comparing the deviation between the actual injection amount and the model injection amount.

[0031] Preferably, the control device of the closed-loop adjustment component is a proportional-integral control device, and the adjustment formula is:

[0032] ;

[0033] In the formula, Q cmd is the expected command injection amount, Q model is the injection amount calculated by the model, DeltaQ is the deviation between the command injection amount and the model injection amount, ET' is the energizing time of the adjusted fuel injector, ET is the energizing time of the unadjusted fuel injector, Kp is the proportional coefficient, and its value range is 0.5 to 2.0, and Ki is the integral coefficient, and its value range is 0.01 to 0.1.

[0034] A HPDI dual-fuel injection control method based on rail pressure signal and neural network, the control method comprising the following steps:

[0035] Step one, collecting real-time rail pressure signals of the engine high-pressure common rail system;

[0036] Step two, pre-processing the collected rail pressure signals, performing transform processing of time domain signals to frequency domain signals to obtain frequency domain features, and extracting frequency band energy E in a specific frequency range;

[0037] Step three, inputting the processed frequency band energy into a preset neural network calculation model, after processing through the preset neural network structure, outputting the model injection amount through the calculation mode of the output part, and the calculation mode of the output part includes coefficients that can be updated online;

[0038] Step four, calculating the deviation between the instruction injection amount and the model injection amount, and dynamically adjusting the energizing time of the fuel injector through the control device according to the deviation to correct the fuel injection amount.

[0039] Preferably, in step one, a high-precision pressure sensor is used to collect rail pressure signals, 10,000 data points per second are acquired during collection, the collection range covers the rail pressure interval of 0 to 200 MPa, and the working condition parameters including engine speed and load are recorded synchronously during the collection process. The real-time rail pressure signals collected cover the rail pressure data of the engine under different working conditions, including idle speed, acceleration, constant speed and deceleration. The amount of data collected under each working condition is not less than 20% of the total data amount, and the collected data is backed up and updated regularly.

[0040] Preferably, in step two, the pre-processing includes removing pulse noise and high-frequency interference in the signal, and performing smoothing filtering, and for every 1024 consecutive data points, transform processing of time domain signals to frequency domain signals is performed, these data points correspond to a time window of 102.4 milliseconds, complex spectrum values corresponding to each frequency are obtained, the specific frequency range extracted is 100Hz to 500Hz, and abnormal spectrum values are removed during the extraction process; the calculation method of the transform processing of time domain signals to frequency domain signals is:

[0041] ;

[0042] In the formula, P(f) is the complex spectrum value corresponding to the frequency f, p(k) is the amplitude of the discretized time domain signal at the kth sampling point, N is the total number of signal sampling points, e is the natural constant, j is the imaginary unit, f is the target frequency component, and k is the serial number of the sampling point;

[0043] A fast algorithm is used in the calculation process to shorten the processing time; the calculation method of the frequency band energy E in the range of 100Hz to 500Hz is .

[0044] Preferably, the preset neural network structure of the neural network calculation model in step three comprises two layers of full connection hidden parts, each layer of full connection hidden part comprises 16 processing units, adopts linear rectification operation as the activation mode, the weight matrix is 1 row 16 columns and 16 rows 16 columns respectively, and the initial value of the weight matrix is obtained through offline training; when offline training, a sample set covering 1000 hours of different working condition operation data is used for model training;

[0045] The calculation method of the output part is:

[0046] Q model =k×ln(E+c)+b;

[0047] In the formula, Q model is the model injection amount, E is the frequency band energy, k, c and b are three constant coefficients; the coefficients that can be updated online are three, which are updated through online learning to adapt to different wear stages of the fuel injector; in the initial coefficients of the new fuel injector, k is 0.18, c is 12.7, and b is 5.2;

[0048] In the coefficients of the worn fuel injector, k is 0.22, c is 14.1, and b is 6.8; when online learning, the coefficients are calibrated and updated according to the deviation between the actual injection amount and the model injection amount every 100 hours of operation.

[0049] Preferably, the control device in step four is a proportional integral control device, and the adjustment formula is:

[0050] ;

[0051] In the formula, Q cmd is the expected command injection amount, Q model is the injection amount calculated by the model, DeltaQ is the deviation between the command injection amount and the model injection amount, ET' is the energizing time of the adjusted fuel injector, ET is the energizing time of the unadjusted fuel injector, Kp is the proportional coefficient, and the value range is 0.5 to 2.0, Ki is the integral coefficient, and the value range is 0.01 to 0.1; the corrected real-time monitoring injection amount is fed back to the next step of calculation; by dynamically adjusting the energizing time of the fuel injector, the self-adaptive compensation of the injection amount of each cylinder in the whole life cycle of the fuel injector is carried out.

[0052] (Three) beneficial effects

[0053] 1. Employing a high-precision pressure sensor, it acquires 10,000 data points per second, covering the rail pressure range of 0 to 200 MPa, providing accurate raw data for subsequent processing; the signal processing and analysis component extracts key frequency band energy through preprocessing and frequency domain conversion, providing reliable input for neural network calculation; the neural network calculation component, with its preset structure and online-updable coefficients, can accurately calculate the model injection quantity; the closed-loop adjustment component dynamically adjusts the injector energizing time through proportional-integral control, ultimately maintaining the accuracy of dual-fuel proportional control, significantly improving control accuracy compared to traditional methods, and ensuring stable engine output.

[0054] 2. New and worn injectors have different initial coefficients, and the system is calibrated and updated every 100 hours of operation based on actual deviations, effectively adapting to different wear stages of the injectors. Simultaneously, the collected rail pressure signals cover various operating conditions, including engine idling and acceleration, with each condition accounting for no less than 20% of the total data. This allows the neural network model to fully learn the characteristics of different operating conditions, enabling the system to operate stably in complex scenarios such as truck climbing hills and heavy-duty construction machinery operations. This effectively solves the problem of poor adaptability of traditional methods to complex operating conditions and injector dynamic states. Furthermore, precise injection quantity control avoids fuel waste caused by injection quantity deviations, improving fuel efficiency and reducing operating costs. At the same time, nitrogen oxide emission fluctuations are controlled within ±2.9%, reducing harmful gas emissions, meeting environmental protection requirements, and contributing to efficient and environmentally friendly engine operation in various scenarios. Attached Figure Description

[0055] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0056] Figure 1 This is an overall structural diagram of an embodiment of the present invention;

[0057] Figure 2 This is a control system diagram in an embodiment of the present invention;

[0058] Figure 3 This is a flowchart of the HPDI dual-fuel injection control method of the present invention. Detailed Implementation

[0059] The embodiment of the application provides a HPDI dual-fuel injection control system and method based on a rail pressure signal and a neural network, solves the technical problem that a traditional injection control method is difficult to balance control precision and adaptability, adopts a high-precision pressure sensor, acquires 10,000 data points per second and covers a 0-200MPa rail pressure interval, provides accurate original data for subsequent processing, a signal processing and analysis component extracts key band energy through preprocessing, frequency domain conversion and other operations, provides reliable input for neural network calculation, a neural network calculation component can accurately calculate model injection amount by virtue of a preset structure and online-updatable coefficients, and a closed-loop adjustment component dynamically adjusts injector energizing time through proportional integral control, finally realizes dual-fuel proportional control precision maintenance, greatly improves control precision compared with a traditional method, and ensures stable engine output.

[0060] Embodiment: The technical solution in the embodiment of the application solves the technical problem that a traditional injection control method is difficult to balance control precision and adaptability, and the general idea is as follows:

[0061] In view of the problems in the prior art, the application provides a HPDI dual-fuel injection control system based on a rail pressure signal and a neural network, the control system is composed of a rail pressure signal acquisition component, a signal processing and analysis component, a neural network calculation component and a closed-loop adjustment component, the components are connected through data transmission lines, and cooperatively complete a control task of injection amount; as shown in the system overall architecture Figure 1 The rail pressure signal acquisition component is an input link of the system and is responsible for acquiring original rail pressure signals; the signal processing and analysis component processes original signals and extracts useful features; the neural network calculation component calculates model injection amount according to the extracted features; and the closed-loop adjustment component adjusts according to the deviation between model injection amount and instruction injection amount, forms a complete closed-loop control system and ensures that injection amount always meets expectations.

[0062] Detailed design and working principle of each component:

[0063] The rail pressure signal acquisition component adopts a high-precision pressure sensor, the sensor has a wide measurement range, that is, 0-200MPa, and can quickly respond; the working principle is that a sensitive element senses pressure changes in a high-pressure common rail, converts pressure signals into electrical signals and outputs; when collecting, 10,000 data points are acquired per second, the rail pressure signals of the engine high-pressure common rail system can be collected in real time and accurately, and original data are provided for subsequent processing; such a high sampling frequency can capture subtle changes in the rail pressure signal, and the changes are often closely related to injection amount and injector state.

[0064] Signal processing and analysis component: connected with the rail pressure signal acquisition component, mainly responsible for the pre-processing and frequency domain analysis of the collected rail pressure signal; the pre-processing includes removing the pulse noise and high-frequency interference in the signal, performing smoothing filtering and temperature compensation; the pulse noise and high-frequency interference mainly come from the noise of the sensor itself and external electromagnetic interference, which can be effectively removed through filtering algorithm to improve the signal-to-noise ratio of the signal; temperature compensation is because the output of the pressure sensor will be affected by temperature, the working temperature of the sensor is obtained through the temperature sensor, and the signal is corrected by using the pre-calibrated temperature compensation coefficient to eliminate the influence of temperature change on signal measurement.

[0065] After the pre-processing is completed, time domain to frequency domain processing is performed on every 1024 data points, 1024 data points correspond to a time window of 102.4 milliseconds, fast Fourier transform algorithm is adopted to convert time domain signal to frequency domain signal, frequency domain features are obtained, and frequency band energy in the range of 100 to 500 Hz is extracted; time domain signal reflects the change of signal with time, while frequency domain signal can reflect the frequency component of signal; the change of fuel injection quantity will cause the energy of rail pressure signal in a certain frequency range to change, therefore, the energy of the frequency band can be extracted as an effective feature reflecting fuel injection quantity.

[0066] Neural network calculation component: connected with the signal processing and analysis component, the core of which is a preset neural network model; the neural network model contains two fully connected hidden layers, each layer has 16 neurons, adopts linear rectifier function as the activation function, and the weight matrix is 1 row 16 columns and 16 rows 16 columns respectively; the initial value of the weight matrix is obtained through offline training, and the sample set covering 1000 hours of different operating data is used for offline training; the working principle of the neural network is to perform weighted summation on the input signal through the connection weight between neurons, and perform nonlinear transformation through the activation function to realize layer-by-layer extraction and mapping of the input features; the linear rectifier function as the activation function can increase the nonlinear expression ability of the neural network, and effectively alleviate the gradient disappearance problem; the input of the neural network calculation component is the processed frequency band energy, and the output is the model injection quantity.

[0067] Q model =k x ln(E+c)+b;

[0068] In the formula, Q model is the model injection quantity, E is the frequency band energy, k, c and b are three constant coefficients, and the three coefficients can be updated online to adapt to different states of the fuel injector; the natural logarithm function can perform nonlinear transformation on the input frequency band energy, so that the model can better fit the relationship between fuel injection quantity and frequency band energy.

[0069] The closed-loop adjustment component is connected with the neural network calculation component and the engine injector respectively, receives the command injection amount and the model injection amount, calculates the deviation between the two, and then dynamically adjusts the energizing time of the injector according to the deviation through a proportional integral control device, and the adjustment mode is

[0070] ;

[0071] In the formula, Q cmd is the expected command injection amount, Q model is the injection amount calculated by the model, Delta Q is the deviation between the command injection amount and the model injection amount, ET' is the energizing time of the adjusted injector, ET is the energizing time of the unadjusted injector, Kp is the proportional coefficient, the value range is 0.5 to 2.0, Ki is the integral coefficient, the value range is 0.01 to 0.1; the proportional part can quickly respond according to the size of the deviation and reduce the deviation; the integral part can eliminate the steady-state deviation and improve the control accuracy; by adjusting the energizing time of the injector, the opening time of the injector can be changed, so as to realize the accurate control of the injection amount.

[0072] The flow of the HPDI dual-fuel injection control method based on the rail pressure signal and the neural network is as follows:

[0073] Step one, collect the real-time rail pressure signal of the engine high-pressure common rail system;

[0074] Step two, pre-process, frequency domain conversion and frequency band energy extraction are performed on the rail pressure signal;

[0075] Step three, input the frequency band energy into the neural network calculation model to obtain the model injection amount;

[0076] Step four, according to the deviation between the command injection amount and the model injection amount, adjust the energizing time of the injector to correct the injection amount.

[0077] The flow forms a closed-loop control loop, and through continuous collection, processing, calculation and adjustment, real-time control of the injection amount is realized.

[0078] Rail pressure signal acquisition: high-precision pressure sensor is used to collect real-time rail pressure signals of the engine high-pressure common rail system at a collection frequency of 10,000 data points per second, covering a range of 0 to 200 MPa; during the collection process, engine speed, load and other working condition parameters are recorded synchronously, the collected signals cover engine idle, acceleration, uniform speed and deceleration conditions, and the amount of data collected under each condition is not less than 20% of the total data amount, and the collected data is backed up and updated regularly; collecting high-frequency data is to ensure that transient changes in rail pressure signals can be captured, which contain a wealth of injection information; synchronous recording of working condition parameters can provide a comprehensive working condition background for subsequent neural network training and control strategy adjustment, so that the model can better adapt to different working conditions.

[0079] Signal processing and analysis: The collected rail pressure signals are preprocessed, including removing pulse noise, high-frequency interference, smoothing filtering and temperature compensation; after preprocessing, time-to-frequency domain processing is performed on each continuous 1024 data points (corresponding to a 102.4 millisecond time window) to obtain complex spectrum values corresponding to each frequency, and 100 to 500 Hz band energy is extracted, and the band energy is , and the complex spectrum values are checked for validity before calculation; the principle of time-to-frequency domain processing is to use the properties of Fourier transform to convert signals in the time domain that are difficult to analyze into the frequency domain, which facilitates the extraction of features in a specific frequency range and eliminates outliers; the calculation method of the transformation process of converting time domain signals to frequency domain signals is:

[0080] ;

[0081] In the formula, P(f) is the complex spectrum value corresponding to the frequency f, p(k) is the amplitude of the discretized time domain signal at the kth sampling point, N is the total number of signal sampling points, e is the natural constant, j is the imaginary unit, f is the target frequency component, and k is the serial number of the sampling point;

[0082] A fast algorithm is used in the calculation process to shorten the processing time; the calculation method of the band energy E in the frequency range of 100 Hz to 500 Hz is ; the complex spectrum values are checked for validity before calculation; the principle of time-to-frequency domain processing is to use the properties of Fourier transform to convert signals in the time domain that are difficult to analyze into the frequency domain, which facilitates the extraction of features in a specific frequency range; the band energy of 100 to 500 Hz is closely related to the injection amount, and by extracting this energy, the neural network can provide effective input features.

[0083] Neural network calculation: the processed band energy is input into a preset neural network calculation model, which outputs the model injection amount after being processed by two fully connected hidden layers; the calculation process of the neural network is:

[0084] Q model =k×ln(E+c)+b;

[0085] In the formula, Q model is the model injection quantity, E is the frequency band energy, k, c, and b are three constant coefficients; the coefficients that can be updated online are three, which are updated through online learning to adapt to different wear stages of the fuel injector; the initial coefficients of the new fuel injector are k=0.18, c=12.7, and b=5.2; that is, the input frequency band energy is first multiplied by the weight matrix of the first layer of hidden layer and added to the bias term, and after being activated by the linear rectification function, the output of the first layer of hidden layer is obtained; the output of the first layer of hidden layer is multiplied by the weight matrix of the second layer of hidden layer and added to the bias term, and after being activated by the linear rectification function, the output of the second layer of hidden layer is obtained; finally, the output of the second layer of hidden layer is calculated through the output layer to obtain the model injection quantity; the three coefficients can be updated online, and the initial coefficients of the new fuel injector are k=0.18, c=12.7, and d=5.2, and the coefficients of the worn fuel injector are k=0.22, c=14.1, and d=6.8; the coefficients are calibrated and updated every 100 hours according to the deviation of the actual injection quantity and the model injection quantity; the principle of online learning is to compare the deviation of the actual injection quantity and the model injection quantity, adjust the coefficients by using gradient descent algorithm and the like, make the model injection quantity closer to the actual demand, and thus adapt to the wear change of the fuel injector.

[0086] Closed-loop adjustment: calculate the deviation between the command injection quantity and the model injection quantity, dynamically adjust the power-on time of the fuel injector according to the deviation by a proportional-integral control device, correct the real-time monitoring of the injection quantity change and feed back to the next calculation; the role of the proportional part is to adjust the power-on time immediately according to the size of the deviation, the larger the deviation, the larger the adjustment, so that the system responds to the deviation quickly; the integral part is to accumulate the deviation, when there is a steady-state deviation, the integral term will gradually increase, and the power-on time adjustment will be promoted until the deviation is eliminated; in this way, the deviation of the injection quantity can be compensated in real time, and the actual injection quantity is consistent with the command injection quantity.

[0087] The experiment is carried out on an engine test bench, and the experimental object is a HPDI dual-fuel engine; the experimental equipment includes an engine test bench, a data acquisition system, an oil consumption instrument, an emission analyzer and the like; the engine test bench can simulate different load and speed conditions to provide a stable operating environment for the experiment; the data acquisition system is used to collect parameters such as rail pressure signal, engine speed, load and the like, and the sampling frequency is consistent with the system acquisition component; the oil consumption instrument is used to accurately measure the injection quantity, and provides a basis for verifying the control precision; the emission analyzer is used to measure the concentration of harmful gases such as nitrogen oxides in the engine exhaust, and to evaluate the environmental protection performance of the system; the experimental environment temperature is controlled at about 25℃, and the humidity is controlled at about 50%, so as to reduce the influence of environmental factors on the experimental results.

[0088] Experimental scheme design: control precision experiment under different working conditions: under different working conditions such as engine idle, acceleration, uniform speed, deceleration, etc., the control method of the present application and the traditional control method are used respectively, the control precision of dual fuel ratio is measured, and the control effects of the two methods are compared; each working condition is continuously operated for 30 minutes, and the dual fuel ratio data is recorded every 1 minute, and the control precision is evaluated by calculating the deviation of the data.

[0089] Adaptability experiment under the condition of injector wear: select new injectors and injectors with different wear degrees, accelerate wear through bench test, simulate wear state under different use time, under the same working condition, use the control method of the present application, measure the control precision of fuel injection quantity and the performance parameters of engine such as power, torque, emission, etc., verify the adaptability of the system to the wear state of the injector; run for 1 hour under each wear state, record the related parameters and analyze.

[0090] Stability experiment under different environmental conditions: under different environmental temperature (-40℃ to 80℃) and altitude (0 to 5000 meters), the control method of the present application is used to measure the control precision and stability of the system, and the working performance of the system under different environmental conditions is verified; different temperature and altitude environmental conditions are realized through environmental simulation cabin, run for 2 hours under each environmental condition, record the change of control precision.

[0091] Finally, it should be noted that: obviously, the above examples are only examples for clearly illustrating the present application, and are not limited to the implementation mode. For ordinary skilled in the art, on the basis of the above description, other different forms of changes or variations can also be made. Here, it is not necessary and impossible to enumerate all the implementation modes. The obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. An HPDI dual-fuel injection control system based on rail pressure signal and neural network, characterized in that, The control system includes: Rail pressure signal acquisition unit, used to acquire real-time rail pressure signals of the engine's high-pressure common rail system; The signal processing and analysis unit is connected to the rail pressure signal acquisition unit. It preprocesses the acquired rail pressure signal, performs transformation processing to convert the time domain signal into a frequency domain signal to obtain frequency domain characteristics, and extracts the frequency band energy E within a specific frequency range. The neural network computing component is connected to the signal processing and analysis component. Its input part receives the processed frequency band energy, and after processing by the preset neural network structure, the output part outputs the model jet amount. The calculation method of the output part includes coefficients that can be learned and updated online. The closed-loop adjustment component is connected to the neural network calculation component and the engine injector respectively. It receives the command injection quantity and the model injection quantity, calculates the deviation between the two, and dynamically adjusts the energizing time of the injector according to the deviation through the control device.

2. The HPDI dual-fuel injection control system based on rail pressure signal and neural network according to claim 1, characterized in that, The rail pressure signal acquisition component uses a high-precision pressure sensor, which acquires 10,000 data points per second during acquisition, and the acquisition range covers the rail pressure range from 0 to 200 MPa. The preprocessing of the signal processing and analysis unit includes removing impulse noise and high-frequency interference from the signal and performing smoothing filtering. For every 1024 consecutive data points, a transformation process is performed to convert the time-domain signal into a frequency-domain signal. These data points correspond to a time window with a duration of 102.4ms, and the specific frequency range extracted is 100Hz to 500Hz. The signal processing and analysis unit's preprocessing of the rail pressure signal also includes temperature compensation, which corrects the acquired signal based on the sensor's operating temperature.

3. The HPDI dual-fuel injection control system based on rail pressure signal and neural network according to claim 1, characterized in that, The preset neural network structure of the neural network computing component includes two fully connected hidden layers, each containing 16 processing units. Linear rectified operation is used as the activation method, and the weight matrices are 1x16 and 16x16 respectively. The initial values ​​of the weight matrices are obtained through offline training. The calculation formula for the output part is: Q model =k×ln(E+c)+b; In the formula, Q model Let E be the model injection quantity, E be the frequency band energy, and k, c, and b be three constant coefficients.

4. The HPDI dual-fuel injection control system based on rail pressure signal and neural network according to claim 1, characterized in that, The neural network computing component has three coefficients that can be learned and updated online. These coefficients are updated online to adapt to different wear stages of the injector. In the initial coefficients of the new injector, k is 0.18, b is 12.7, and c is 5.

2. In the coefficients for wear on fuel injectors, k is 0.22, b is 14.1, and c is 6.

8. During online learning, the coefficients are calibrated and updated every 100 hours of operation. During the update process, the coefficients are corrected by comparing the deviation between the actual fuel injection quantity and the model injection quantity.

5. The HPDI dual-fuel injection control system based on rail pressure signal and neural network according to claim 1, characterized in that, The control device for the closed-loop adjustment component is a proportional-integral control device, and the adjustment formula is: ； In the formula, Q cmd Q is the desired command injection quantity. model DeltaQ is the injection quantity calculated by the model, DeltaQ is the deviation between the commanded injection quantity and the model injection quantity, ET′ is the energizing time of the injector after adjustment, ET is the energizing time of the injector before adjustment, Kp is the proportional coefficient, with a value range of 0.5 to 2.0, and Ki is the integral coefficient, with a value range of 0.01 to 0.

1.

6. A dual-fuel injection control method for HPDI based on rail pressure signal and neural network, characterized in that, Includes the following steps: Step 1: Acquire the real-time rail pressure signal of the engine's high-pressure common rail system; Step 2: Preprocess the acquired rail pressure signal by converting the time-domain signal into a frequency-domain signal to obtain frequency domain characteristics and extracting the frequency band energy E within a specific frequency range. Step 3: Input the processed frequency band energy into the preset neural network calculation model. After processing by the preset neural network structure, the model jet amount is output through the calculation method of the output part. The calculation method of the output part includes coefficients that can be learned and updated online. Step 4: Calculate the deviation between the commanded injection quantity and the model injection quantity, and dynamically adjust the energizing time of the injector according to the deviation through the control device to correct the injection quantity.

7. The HPDI dual-fuel injection control method based on rail pressure signal and neural network according to claim 6, characterized in that, In step one, a high-precision pressure sensor is used to collect rail pressure signals. During the collection, 10,000 data points are acquired per second, covering the rail pressure range from 0 to 200 MPa. During the collection process, operating parameters including engine speed and load are recorded simultaneously. The collected real-time rail pressure signals cover rail pressure data of the engine under different operating conditions, including idling, acceleration, constant speed and deceleration. The amount of data collected under each operating condition is not less than 20% of the total data. The collected data is backed up and updated regularly.

8. The HPDI dual-fuel injection control method based on rail pressure signal and neural network according to claim 6, characterized in that, Step two preprocessing includes removing impulse noise and high-frequency interference from the signal and performing smoothing filtering. For every consecutive 1024 data points, a transformation process converting the time-domain signal to a frequency-domain signal is performed. These data points correspond to a time window of 102.4 milliseconds, obtaining the complex spectrum values ​​corresponding to each frequency. The specific frequency range extracted is 100Hz to 500Hz, and abnormal spectrum values ​​are removed during the extraction process. The calculation method for the transformation process converting the time-domain signal to a frequency-domain signal is as follows: ； In the formula, P(f) is the complex spectrum value corresponding to frequency f, p(k) is the amplitude of the discretized time-domain signal at the kth sampling point, N is the total number of sampling points of the signal, e is the natural constant, j is the imaginary unit, f is the target frequency component, and k is the index of the sampling point. A fast algorithm is used to shorten processing time during the calculation; the calculation method for the frequency band energy E in the frequency range of 100Hz to 500Hz is as follows: .

9. The HPDI dual-fuel injection control method based on rail pressure signal and neural network according to claim 6, characterized in that, The neural network structure of the preset neural network computing model in step three includes two fully connected hidden parts. Each fully connected hidden part contains 16 processing units and uses linear rectified operation as the activation method. The weight matrices are 1 row and 16 columns and 16 rows and 16 columns, respectively. The initial values ​​of the weight matrices are obtained through offline training. During offline training, a sample set covering 1000 hours of operation data under different working conditions is used to train the model. The output is calculated as follows: Q model =k×ln(E+c)+b; In the formula, Q model The model injection quantity is E, the frequency band energy is E, and k, c, and b are three constant coefficients. The three coefficients can be updated online to adapt to different wear stages of the injector. In the initial coefficients of the new injector, k is 0.18, c is 12.7, and b is 5.

2. In the coefficients for wear injectors, k is 0.22, c is 14.1, and b is 6.8; during online learning, the coefficients are calibrated and updated every 100 hours based on the deviation between the actual injection quantity and the model injection quantity.

10. The HPDI dual-fuel injection control method based on rail pressure signal and neural network according to claim 6, characterized in that, In step four, the control device is a proportional-integral control device, and the adjustment formula is: ； In the formula, Q cmd Q is the desired command injection quantity. model DeltaQ is the injection quantity calculated by the model, DeltaQ is the deviation between the commanded injection quantity and the model injection quantity, ET′ is the energizing time of the injector after adjustment, ET is the energizing time of the injector before adjustment, Kp is the proportional coefficient, with a value range of 0.5 to 2.0, and Ki is the integral coefficient, with a value range of 0.01 to 0.

1. After correction, the changes in fuel injection quantity are monitored in real time and fed back to the next calculation; by dynamically adjusting the power-on time of the injector, adaptive compensation of the fuel injection quantity of each cylinder is performed throughout the life cycle of the injector.

Images