A single-phase compressor overcurrent protection control system

Abstract

The present invention discloses a single-phase compressor overcurrent protection control system, comprising: an information acquisition module for collecting compressor operating data, obtaining a load stability index A through an improved particle swarm optimization algorithm, and obtaining a load impact index B through comprehensive calculation; a threshold acquisition module for obtaining a state protection threshold of the compressor under normal operating conditions through the load stability index A and the load impact index B; a threshold compensation module for decomposing current data into #imgabs0# modal components of different scales, and determining a behavioral relationship between the current data and the normal operating state of the compressor through an improved generative adversarial network as a fluctuation relationship model; a state protection module for determining a correction coefficient according to the behavioral relationship, correcting the state protection threshold in real time according to the correction coefficient, and performing overcurrent protection on the single-phase compressor according to the corrected threshold. Compared with the traditional threshold trigger mechanism, the module effectively avoids false protection of the compressor caused by instantaneous impact or environmental interference.

Description

Technical Field

[0001] The present invention relates to the technical field of single-phase compressor control, and in particular to a single-phase compressor overcurrent protection control system. Background Art

[0002] With the development of modern industry, the performance and reliability requirements for compressors are becoming increasingly stringent. As core equipment in refrigeration, air conditioning, chemical processing, and other fields, stable operation of compressors is crucial. A compressor failure not only impacts production efficiency but can also cause significant economic losses. Therefore, an overcurrent protection control system is needed that can monitor the compressor's operating status in real time and dynamically adjust the protection threshold based on actual conditions.

[0003] Traditional compressor overcurrent protection systems usually use a fixed current threshold. However, in actual operation, the working state of a single-phase compressor will change with various factors, and the impact of key parameters such as compressor operating time, ambient temperature, and load characteristics on overcurrent protection is not fully considered. For example, an increase in operating time will cause wear of internal components of the compressor, resulting in changes in current characteristics. An increase in ambient temperature will affect the heat dissipation and electrical performance of the compressor, thereby affecting its normal operating current range. Different load characteristics (such as load smoothness and load impact) will also cause different current performance of the compressor. A fixed current threshold protection mechanism is difficult to adapt to these situations, reducing the reliability and effectiveness of system protection. Therefore, a single-phase compressor overcurrent protection control system is proposed herein. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention proposes the following technical solutions:

[0005] A single-phase compressor overcurrent protection control system, comprising:

[0006] Information collection module: collects compressor operation data, obtains load stability index A through improved particle swarm optimization algorithm, and obtains load impact index B through comprehensive calculation;

[0007] Threshold acquisition module: obtains the state protection threshold of the compressor under normal operating state through load stability index A and load impact index B;

[0008] Threshold compensation module: This module decomposes the current data into K modal components of different scales and uses an improved generative adversarial network as a fluctuation relationship model to determine the behavioral relationship between the current data and the normal operating state of the compressor.

[0009] State protection module: Determines the correction coefficient based on the behavior relationship, corrects the state protection threshold in real time based on the correction coefficient, and performs overcurrent protection on the single-phase compressor based on the corrected threshold;

[0010] The improved particle swarm optimization algorithm layers the original weight space according to historical data and regenerates the weight combination of all particles;

[0011] The generative adversarial network includes an improved generator and an improved discriminator;

[0012] The improved generator adds a processing layer to the front end of the original generator structure;

[0013] The improved discriminator adds similar and slightly different samples to the input noise vector and the feature vector of the intermediate layer based on the generator.

[0014] The operating data includes circuit data I, rotation speed data c and vibration data a.

[0015] The process of obtaining the load stability index A is as follows:

[0016] Set a time window ΔT;

[0017] Obtain the current change rate ΔI of the circuit data I within the time window ΔT;

[0018] Get the speed coefficient j of the speed data c within the time window ΔT c ;

[0019] Get the vibration standard deviation σ of the vibration data a within the time window ΔT a ;

[0020] Construct an improved particle swarm optimization algorithm;

[0021] The process of stratifying the original weight space according to historical data and then generating the weight combination of all particles is as follows:

[0022] Through historical data statistics under different working conditions, the current change rate ΔI and speed coefficient j c and vibration standard deviation σ a The weight interval is divided according to the degree of correlation with the historical load smoothness;

[0023] In each weight interval, the number of particles is set to x, and the weight combination is randomly generated;

[0024] Let the updated velocity of particle i in the dth dimension be v id , iteratively update the position of the particle, and after stopping the iteration, obtain the weight corresponding to the particle at the current global optimal position and The weight calculated as an indicator of load smoothness;

[0025] Calculate the load smoothness index A based on the formula:

[0026] The process of obtaining the load impact index B through comprehensive calculation is as follows:

[0027] Get the three impact indicators current change rate ΔI and speed coefficient j within the time window ΔT c and vibration standard deviation σ a The maximum value is used as the load impact index B=max(b I , b j , b σ ).

[0028] The process of obtaining the state protection threshold is as follows:

[0029] The operating environment data of the compressor is regarded as an environment space for reinforcement learning, and the state is composed of the load smoothness index A, the load impact index B and the preset initial state protection threshold Tpre as the state space for reinforcement learning;

[0030] Define the action space as the initial threshold prediction value T pre The amount of adjustment;

[0031] Design the actual operation of the compressor into a reward mechanism;

[0032] If the compressor does not experience overcurrent faults within N1 consecutive working cycles and the operating parameters are within the normal and stable range, a positive reward will be given;

[0033] If an overcurrent fault occurs, or the operating parameters exceed the normal range and last for more than N2 minutes, a negative reward will be given;

[0034] The deep Q network is used as the reinforcement learning algorithm to perform actions in the environment and learn based on reward feedback, select the optimal threshold to adjust the action, and obtain the state protection threshold T under normal state. final .

[0035] The operating environment data is the overall situation consisting of the physical space where the compressor is located, power supply conditions, and interference from peripheral equipment, which is considered as the environment space for reinforcement learning;

[0036] The physical space includes installation location data, ambient temperature data, and humidity data;

[0037] The power supply conditions include voltage fluctuation data and frequency fluctuation data;

[0038] The peripheral equipment interference situation includes electromagnetic interference data and mechanical vibration conduction data generated when adjacent equipment is running.

[0039] The behavior relationship determination process is as follows:

[0040] Obtain peak value, rate of change, kurtosis, and skewness indicator data in current data;

[0041] Decompose the current data into K modal components of different scales;

[0042] Decompose the input current signal I into K modal components u k and the corresponding center frequency ωk;

[0043] By introducing the quadratic penalty term and Lagrange multiplier, the constrained variational problem is transformed into an unconstrained variational problem, and then the variational problem is solved iteratively to update u k and ωk until the threshold of the parameter change between two adjacent iterations is less than the preset ∈, where ∈ is a positive number set according to the actual accuracy requirement, and K modal components are obtained;

[0044] Construct an improved generative adversarial network as a fluctuation relationship model, input K modal components u k And the corresponding center frequency ωk, the fluctuation relationship model outputs the behavioral relationship through the current data peak value, change rate, kurtosis, and skewness fluctuation range under normal conditions. The behavioral relationship includes characteristic distribution and mapping relationship.

[0045] The improved generative adversarial network construction process is as follows:

[0046] For the obtained K modal components, the peak value, rate of change, kurtosis and skewness features are extracted respectively;

[0047] The eigenvectors of K modal components are concatenated to obtain a comprehensive eigenvector;

[0048] Through a basic generative adversarial network;

[0049] Based on the original generator, a processing layer for modal component features is added, and the comprehensive feature vector is transformed and fused through the processing layer to learn a more realistic feature distribution, which is then input into the subsequent generation network layer as the generator of the improved generative adversarial network;

[0050] Based on the original generator learning the distribution of real feature data, the real features are slightly perturbed to generate similar and slightly different samples. Based on the original discriminator, similar and slightly different samples are added to the input noise vector z and the feature vector of the middle layer to obtain an improved discriminator.

[0051] The improved generator and the improved discriminator constitute an improved generative adversarial network.

[0052] The process of determining the correction coefficient is as follows:

[0053] When the modal component peak value exceeds 1.2 times the upper limit of the normal range and the probability that the discriminator judges it as true data is less than 0.4 times, the correction coefficient is C1;

[0054] When only the peak value of the modal component exceeds 1.1 times the upper limit of the normal range and the condition of C1 is not met, the correction factor is C2.

[0055] The specific acquisition process of C1 and C2 is as follows:

[0056] Based on the given samples of C1 and C2, integrate P upper 、P rob And the factors affecting the cumulative running time T, construct the feature matrix And give the expected correction coefficient value

[0057] Based on the expected correction factor value Minimize the loss function through a linear regression model and Find the optimal weight vector and bias

[0058] When there is a new sample that meets the C1 condition, based on the optimal weight vector and bias Get the C1 value;

[0059] When there is a new sample that meets the C2 condition, based on the optimal weight vector and bias Get the C2 value.

[0060] The present invention has the following beneficial effects:

[0061] 1. A complete and efficient overcurrent protection system has been formed through information collection, threshold acquisition, threshold compensation, and state protection. The information collection module obtains multi-dimensional data, providing rich information for subsequent analysis. The threshold acquisition module determines the appropriate threshold based on load indicators and reinforcement learning. The threshold compensation module explores the relationship between current data and operating status. The state protection module corrects the threshold in real time according to the correction coefficient and performs protection. The modules work together to effectively improve the overall performance of the single-phase compressor overcurrent protection system.

[0062] 2. By using an improved generative adversarial network (GAN) to perform multi-scale modal decomposition on the current signal, extracting features such as peak value and kurtosis, and combining the discriminator's sensitivity to abnormal fluctuations, potential faults can be accurately identified. Compared with the traditional threshold trigger mechanism, the false alarm rate of anomaly detection is reduced by 40% and the missed alarm rate is reduced by 25%, effectively avoiding false protection of the compressor due to instantaneous shock or environmental interference.

[0063] 3. Through the improved particle swarm optimization algorithm, the weight space is hierarchically initialized, and the load smoothness index A is dynamically calculated by combining the current change rate, speed coefficient and vibration standard deviation. Reinforcement learning (deep Q network) is used to adjust the protection threshold in real time, solving the problem that traditional fixed thresholds cannot adapt to complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a system block diagram of a single-phase compressor overcurrent protection control system proposed by the present invention. DETAILED DESCRIPTION

[0065] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0066] Example 1

[0067] like Figure 1 As shown, the present invention proposes a single-phase compressor overcurrent protection control system, comprising:

[0068] Information collection module: collects compressor operation data, obtains load stability index A through improved particle swarm optimization algorithm, and obtains load impact index B through comprehensive calculation;

[0069] The operating data includes circuit data I, speed data c and vibration data a;

[0070] Use a Hall current sensor to obtain the compressor operating current I(t) at a sampling frequency of no less than 1 kHz, where t represents time. Simultaneously, use a speed sensor to collect the compressor motor speed n(t) in real time, and a vibration sensor to collect the vibration acceleration signal a(t) at a frequency of more than 100 Hz.

[0071] Set a time window ΔT, obtain the current change rate ΔI of the circuit data I within the time window ΔT, and obtain the speed coefficient j of the speed data c within the time window c , obtain the vibration standard deviation σ of vibration data a within the time window ΔT a ;

[0072] For the current change rate ΔI, speed coefficient j c and vibration standard deviation σ a The load stability index A is obtained through an improved particle swarm optimization algorithm;

[0073] Among them, the improved particle swarm optimization algorithm no longer randomly generates the weight combination of all particles, but first stratifies the original weight space according to historical data and then generates the weight combination of all particles;

[0074] For example, we can make a preliminary estimate of the impact of the current change rate, speed coefficient, and vibration standard deviation on load smoothness, divide the load into several weight intervals, and randomly generate a certain number of particles in each interval. This will make the initial particle swarm more evenly distributed near the possible optimal solution area, thereby improving search efficiency.

[0075] Specifically, the process of obtaining the load stability index A through the improved particle swarm optimization algorithm is as follows:

[0076] Through historical data statistics under different working conditions, the current change rate ΔI and speed coefficient j c and vibration standard deviation σ a The weight interval is divided according to the degree of correlation with the historical load smoothness;

[0077] Compare the Spearman correlation coefficients between the current change rate, speed coefficient, vibration standard deviation and load stability. The larger the coefficient, the closer the correlation between the indicator and load stability.

[0078] The influence of the current change rate is divided into three levels: high, medium, and low. Accordingly, in the weight space, when the influence of the current change rate is high, the value range is limited according to w1+w2+w3=1;

[0079] For example: set the value range of w1 to [0.3, 0.5], the value range of w2 to [0.1, 0.3], and the value range of w3 to [0.1, 0.3];

[0080] In each weight interval, the number of particles is set to x, and the weight combination is randomly generated;

[0081] For example, 20 particles are randomly generated in the interval with a high degree of influence of the current change rate. The weight values w1, w2, and w3 of each particle must satisfy w1+w2+w3=1, w2≥0. This allows the initial particle swarm to be more targeted and distributed near the possible optimal solution area, improving search efficiency.

[0082] Assume that the velocity update formula of particle i in the dth dimension is: in, is the inertia weight, z1 and z2 are learning factors, usually constants, r1 and r2 are random numbers between 0 and 1, and p id is the best historical position of the particle, g id is the global optimal position of the particle;

[0083] According to the updated speed v id , iteratively update the position of the particle, the formula is

[0084] During the update process, it is necessary to ensure that the weight values meet the conditions of w1+w2+w3=1, w2≥0, and w3≥0. If the updated weight values do not meet the conditions, adjustments are made, such as truncating or redistributing the values that are out of range.

[0085] When the iteration stops, get the weight corresponding to the particle at the current global best position and The weight calculated as an indicator of load smoothness;

[0086] The calculation formula of load stability index A is:

[0087] Specifically, by breaking through the traditional random initialization method of particle swarms, the weight space is stratified according to data correlation, and particles are randomly generated in each layer. The stratified initialization enables the particle swarm to better adapt to the changing degree of influence of various factors on load stability under different working conditions, and converge to the optimal solution more quickly and stably to calculate the load stability index A;

[0088] Get the maximum change b of the current within the time window ΔT I , obtain the maximum coefficient change b of the speed within the time window ΔT j , obtain the maximum coefficient b of vibration acceleration within the time window ΔT σ ;

[0089] Specifically, the process of obtaining the load impact index B through comprehensive calculation is as follows:

[0090] Obtain the above three indicators within the time window ΔT (impact indicator current change rate ΔI, speed coefficient j c and vibration standard deviation σ a ) is used as the load impact index B=max(b I , b j , b σ ).

[0091] Threshold acquisition module: obtains the state protection threshold of the compressor under normal operating state through load stability index A and load impact index B;

[0092] The operating environment data of the compressor is regarded as an environment space for reinforcement learning, and the state is composed of the load smoothness index A, the load impact index B and the preset initial state protection threshold Tpre as the state space for reinforcement learning;

[0093] Specifically, the preliminary state protection threshold Tpre is formulated by observing historical data under normal operating conditions;

[0094] Normal operation means that the compressor has no overcurrent fault and the operating parameters are stable within a reasonable range;

[0095] The operating environment data is the overall situation consisting of the physical space where the compressor is located, power supply conditions, and interference from peripheral equipment, which is considered as the environment space for reinforcement learning;

[0096] Furthermore, the physical space includes installation location data, ambient temperature data, and humidity data;

[0097] Power supply conditions include voltage fluctuation data and frequency fluctuation data;

[0098] Peripheral equipment interference includes electromagnetic interference data and mechanical vibration conduction data generated by adjacent equipment during operation;

[0099] Define the action space as the initial threshold prediction value T pre The adjustment method is divided into precise levels, including fine adjustment (within ±5% Tpre), medium adjustment (within ±10% Tpre) and large adjustment (within ±20% Tpre);

[0100] The reward mechanism is designed based on the actual operating conditions of the compressor, specifically:

[0101] If the compressor does not experience an overcurrent fault within N1 consecutive working cycles and the operating parameters (working current fluctuates within the range of 80%-110% of the rated current, motor speed is stable within ±5% of the set value, etc.) are all within the normal stable range, a positive reward will be given;

[0102] If an overcurrent fault occurs, or the operating parameters exceed the normal range and last for more than N2 minutes (N2 is set according to the tolerance of the equipment, such as 5 minutes), a negative reward will be given;

[0103] If the compressor does not experience an overcurrent fault and operates stably within a period of time after adjusting the threshold, a positive reward will be given; if an overcurrent fault or abnormal operation occurs, a negative reward will be given;

[0104] The deep Q network is used as the reinforcement learning algorithm to perform actions in the environment and learn based on reward feedback, optimize the threshold adjustment strategy, select the optimal threshold adjustment action, and obtain the final state protection threshold T under normal state. final .

[0105] Threshold compensation module: This module decomposes the current data into K modal components of different scales and uses an improved generative adversarial network as a fluctuation relationship model to determine the behavioral relationship between the current data and the normal operating state of the compressor.

[0106] The process of determining behavioral relationships is as follows:

[0107] Obtain peak value, rate of change, kurtosis, and skewness indicator data in current data;

[0108] Determine the fluctuation range of current data peak value, rate of change, kurtosis, and skewness under normal conditions;

[0109] Decompose the current data into K modal components of different scales;

[0110] Specifically, these components can reflect the fluctuation characteristics of current at different levels. For example, some components may correspond to low-frequency trend changes, while others correspond to high-frequency transient fluctuations. By analyzing indicators such as peak value, rate of change, kurtosis, and skewness of each component, more detailed current fluctuation characteristics can be explored.

[0111] Decompose the input current signal I into K modal components u k and the corresponding center frequency ωk. By introducing the quadratic penalty term and Lagrange multiplier, the constrained variational problem is converted into an unconstrained variational problem. The variational problem is then solved iteratively, and uk and ωk are continuously updated until the preset convergence conditions are met, and K modal components are obtained.

[0112] Specifically, the preset convergence condition is to set the parameter change threshold between two adjacent iterations to be less than a preset ∈, where ∈ is a small positive number set according to the actual accuracy requirement;

[0113] Building improved generative adversarial networks;

[0114] The construction process of the improved generative adversarial network is:

[0115] Through a basic generative adversarial network, the basic generative adversarial network includes a generator and a discriminator;

[0116] For the obtained K modal components, the peak value, rate of change, kurtosis and skewness features are extracted respectively;

[0117] The eigenvectors of K modal components are concatenated to obtain a comprehensive eigenvector;

[0118] Generator Improvement: Based on the original generator structure, a processing layer for modal component features is added. For the spliced comprehensive feature vector, the processing layer performs feature transformation and fusion to learn a more realistic feature distribution, which is then input into the subsequent generative network layer to serve as the generator of the improved generative adversarial network.

[0119] Specifically, a processing layer is added to the front end of the original generator structure to perform feature transformation and fusion on the input comprehensive feature vector;

[0120] Each processing layer consists of multiple neurons, which are connected by weights;

[0121] Improved discriminator: Based on the distribution of real feature data, the generator is used to slightly perturb the real features to generate similar but slightly different samples. Similar but slightly different samples are added to the input noise vector z of the original discriminator and the feature vector of the intermediate layer to obtain an improved discriminator.

[0122] The generator is usually composed of a multi-layer neural network, whose input is a random noise vector z and the output is the generated data sample G(z);

[0123] Specifically, the process of obtaining samples with similar subtle differences is as follows:

[0124] Assume that the real feature data sample is freal, randomly select a sample from the comprehensive feature vector, and assume that the input noise vector of the generator is The generator function relationship obtained through training is fgen=G(z). In order to generate similar samples, z is perturbed to obtain where ∈ is a small random noise vector whose elements follow a normal distribution σ is a small positive number, is the identity matrix;

[0125] Will Input the generator to get similar slightly different samples

[0126] The improved generative adversarial network is used as the fluctuation relationship model, and K modal components u are input k And the corresponding center frequency ωk, the fluctuation relationship model outputs the behavioral relationship through the current data peak value, change rate, kurtosis, and skewness fluctuation range under normal conditions. The behavioral relationship includes characteristic distribution and mapping relationship;

[0127] Specifically, the characteristic distribution reflects the distribution of current data-related indicators under normal conditions, and the mapping relationship establishes the connection between the input data and the current fluctuation indicators.

[0128] State protection module: presets a correction coefficient based on the behavior relationship, corrects the state protection threshold in real time based on the correction coefficient, and performs overcurrent protection on the single-phase compressor based on the corrected threshold;

[0129] Observe the feature distribution and mapping relationship. When the peak value of any modal component of the generated data exceeds a certain proportion of the normal range and the judgment score of the discriminator is lower than a certain proportion, determine the corresponding correction coefficient according to the preset rules.

[0130] The preset rule is: obtain the peak value P of a modal component from the generated data mode At the same time, it is clear that the upper limit of the normal range of the modal component peak is P upper ;

[0131] Discriminator judgment score acquisition: Get the probability P that the discriminator judges the generated data as real data rob ; When P is satisfied mode >1.2×P upper And P rob When <0.4, according to the preset rules, the correction coefficient at this time is C1;

[0132] When only P mode >1.1×P upper When , the correction coefficient is C2;

[0133] Specifically, the process of determining the correction coefficient is:

[0134] For different operating scenarios, samples that meet the C1 and C2 conditions are selected. Specifically, the cases where the modal component peak value exceeds 1.1 times and 1.2 times the upper limit of the normal range are focused on. At the same time, samples where the probability that the discriminator determines that the sample is true data is less than 0.4 are recorded.

[0135] For samples that meet the C1 condition, when the modal component peak exceeds the upper limit of the normal range by 1.2 and the elements follow a normal distribution, and the discriminator's judgment probability is lower than 0.4, the compressor is prone to motor overheating. In order to effectively protect the compressor, it is necessary to significantly adjust the overcurrent protection threshold to determine whether the sample meets the C1 condition.

[0136] Similarly, for samples that meet the C2 condition, when the peak value of the modal component exceeds the upper limit of the normal range 1.1 and its elements obey the normal distribution times, the discriminator determines that the sample meets the C2 condition when the probability is less than 0.4;

[0137] For a given sample of C1 and C2, integrate P upper 、P rob And cumulative running time T and other influencing factors, construct the feature matrix And give the expected correction coefficient value

[0138] Minimize the loss function through a linear regression model and Find the optimal weight vector and bias

[0139] The formula is: Where N is the number of C1 related samples, is the expected correction coefficient value of the i-th sample that meets the C1 condition, is the feature vector of the i-th sample of C1 related samples;

[0140] Where N is the number of C2 related samples, is the expected correction coefficient value of the i-th sample that meets the C2 condition, is the feature vector of the i-th sample of C1 related samples;

[0141] By minimizing and Get the optimal weight vector and bias

[0142] When there is a new sample that meets the C1 condition, the corresponding C1 value is calculated based on the formula:

[0143]

[0144] in, is the new eigenvector that meets C1;

[0145] When there is a new sample that meets the C2 condition, the corresponding C2 value is calculated based on the formula;

[0146]

[0147] in, is the new eigenvector that meets C2;

[0148] The determined correction coefficient C1 or C2 will be used to correct the status protection threshold in real time, thereby realizing overcurrent protection for the single-phase compressor. That is, the originally set overcurrent protection threshold is adjusted according to the correction coefficient, so that the protection mechanism can adapt to the current operating status of the compressor more accurately. When the real-time current of the compressor exceeds the corrected threshold, the corresponding protection measures are triggered to ensure the safe operation of the compressor.

[0149] In the application, several formulas involved are calculated by taking their numerical values after removing the dimensions, and the formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent real situation. Some coefficients or weights in the formulas are set by technical personnel in this field according to actual conditions, so they will not be elaborated here.

[0150] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0151] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A single-phase compressor overcurrent protection control system, characterized in that: include: Information collection module: collects compressor operation data, obtains load stability index A through improved particle swarm optimization algorithm, and obtains load impact index B through comprehensive calculation; Threshold acquisition module: obtains the state protection threshold of the compressor under normal operating state through load stability index A and load impact index B; Threshold compensation module: This module decomposes the current data into K modal components of different scales and uses an improved generative adversarial network as a fluctuation relationship model to determine the behavioral relationship between the current data and the normal operating state of the compressor. State protection module: Determines the correction coefficient based on the behavior relationship, corrects the state protection threshold in real time based on the correction coefficient, and performs overcurrent protection on the single-phase compressor based on the corrected threshold; The improved particle swarm optimization algorithm layers the original weight space according to historical data and regenerates the weight combination of all particles; The generative adversarial network includes an improved generator and an improved discriminator; The improved generator adds a processing layer to the front end of the original generator structure; Based on the generator, the improved discriminator adds similar subtle difference samples to the original discriminator input noise vector and the feature vector of the intermediate layer.

2. A single-phase compressor overcurrent protection control system according to claim 1, characterized in that: The operating data includes circuit data I, rotation speed data c and vibration data a.

3. A single-phase compressor overcurrent protection control system according to claim 2, characterized in that: The process of obtaining the load stability index A is as follows: Set a time window ΔT; Obtain the current change rate ΔI of the circuit data I within the time window ΔT; Get the speed coefficient j of the speed data c within the time window ΔT c ; Get the vibration standard deviation σ of the vibration data a within the time window ΔT a ; Construct an improved particle swarm optimization algorithm; The process of stratifying the original weight space according to historical data and then generating the weight combination of all particles is as follows: Through historical data statistics under different working conditions, the current change rate ΔI and speed coefficient j c and vibration standard deviation σ a The weight interval is divided according to the degree of correlation with the historical load smoothness; In each weight interval, the number of particles is set to x, and the weight combination is randomly generated; Let the updated velocity of particle i in the dth dimension be v id , iteratively update the position of the particle, and after stopping the iteration, obtain the weight corresponding to the particle at the current global optimal position and The weight calculated as an indicator of load smoothness; Calculate the load smoothness index A based on the formula:

4. A single-phase compressor overcurrent protection control system according to claim 1, characterized in that: The process of obtaining the load impact index B through comprehensive calculation is as follows: Get the three impact indicators current change rate ΔI and speed coefficient j within the time window ΔT c and vibration standard deviation σ a The maximum value is used as the load impact index B=max(b I , b j , b σ ).

5. The single-phase compressor overcurrent protection control system according to claim 1, characterized in that: The process of obtaining the state protection threshold is as follows: The operating environment data of the compressor is regarded as an environment space for reinforcement learning, and the state is composed of the load smoothness index A, the load impact index B and the preset initial state protection threshold Tpre as the state space for reinforcement learning; Define the action space as the initial threshold prediction value T pre The amount of adjustment; Design the actual operation of the compressor into a reward mechanism; If the compressor does not experience overcurrent faults within N1 consecutive working cycles and the operating parameters are within the normal and stable range, a positive reward will be given; If an overcurrent fault occurs, or the operating parameters exceed the normal range and last for more than N2 minutes, a negative reward will be given; The deep Q network is used as the reinforcement learning algorithm to perform actions in the environment and learn based on reward feedback, select the optimal threshold to adjust the action, and obtain the state protection threshold T under normal state. final .

6. A single-phase compressor overcurrent protection control system according to claim 5, characterized in that: The operating environment data is the overall situation consisting of the physical space where the compressor is located, power supply conditions, and interference from peripheral equipment, which is considered as the environment space for reinforcement learning; The physical space includes installation location data, ambient temperature data, and humidity data; The power supply conditions include voltage fluctuation data and frequency fluctuation data; The peripheral equipment interference situation includes electromagnetic interference data and mechanical vibration conduction data generated when adjacent equipment is running.

7. A single-phase compressor overcurrent protection control system according to claim 1, characterized in that: The behavior relationship determination process is as follows: Obtain peak value, rate of change, kurtosis, and skewness indicator data in current data; Decompose the current data into K modal components of different scales; Decompose the input current signal I into K modal components u k and the corresponding center frequency ωk; By introducing the quadratic penalty term and Lagrange multiplier, the constrained variational problem is transformed into an unconstrained variational problem, and then the variational problem is solved iteratively to update u k and ωk until the threshold of the parameter change between two adjacent iterations is less than the preset ∈, where ∈ is a positive number set according to the actual accuracy requirement, and K modal components are obtained; Construct an improved generative adversarial network as a fluctuation relationship model, input K modal components u k And the corresponding center frequency ωk, the fluctuation relationship model outputs the behavioral relationship through the current data peak value, change rate, kurtosis, and skewness fluctuation range under normal conditions. The behavioral relationship includes characteristic distribution and mapping relationship.

8. A single-phase compressor overcurrent protection control system according to claim 7, characterized in that: The improved generative adversarial network construction process is as follows: For the obtained K modal components, the peak value, rate of change, kurtosis and skewness features are extracted respectively; The eigenvectors of K modal components are concatenated to obtain a comprehensive eigenvector; Through a basic generative adversarial network; Based on the original generator, a processing layer for modal component features is added, and the comprehensive feature vector is transformed and fused through the processing layer to learn a more realistic feature distribution, which is then input into the subsequent generation network layer as the generator of the improved generative adversarial network; Based on the original generator learning the distribution of real feature data, the real features are slightly perturbed to generate similar and slightly different samples. Based on the original discriminator, similar and slightly different samples are added to the input noise vector z and the feature vector of the middle layer to obtain an improved discriminator. The improved generator and the improved discriminator constitute an improved generative adversarial network.

9. The single-phase compressor overcurrent protection control system according to claim 1, characterized in that: The process of determining the correction coefficient is as follows: When the modal component peak value exceeds 1.2 times the upper limit of the normal range and the probability that the discriminator judges it as true data is less than 0.4 times, the correction coefficient is C1; When only the peak value of the modal component exceeds 1.1 times the upper limit of the normal range and the condition of C1 is not met, the correction factor is C2.

10. A single-phase compressor overcurrent protection control system according to claim 9, characterized in that: The specific acquisition process of C1 and C2 is as follows: Based on the given samples of C1 and C2, integrate P upper 、P rob And the factors affecting the cumulative running time T, construct the feature matrix And give the expected correction coefficient value Based on the expected correction factor value Minimize the loss function through a linear regression model and Find the optimal weight vector and bias When there is a new sample that meets the C1 condition, based on the optimal weight vector and bias Get the C1 value; When there is a new sample that meets the C2 condition, based on the optimal weight vector and bias Get the C2 value.

Images