Cold storage group collaborative energy-saving control system based on multi-agent reinforcement learning

Abstract

The present application relates to the field of cold storage group energy management control, and particularly relates to a cold storage group collaborative energy-saving control system based on multi-agent reinforcement learning. The present application converts the bearing grease deterioration degree index of the evaporator fan bearing vibration characteristics under the normalization of operating conditions and mechanical defect suppression, obtains the prediction value by using the historical degradation trajectory similarity matching, and puts the current and predicted deterioration index into the multi-agent reinforcement learning state and reward, so as to obtain a collaborative energy-saving control system considering the equipment health state and the cold storage group demand constraint. The present application reduces the total power peak and invalid refrigeration energy consumption of the cold storage group, reduces the risk of bearing grease deterioration misjudgment and sudden failure, improves the operation safety and economy of multi-cold storage shared refrigeration, realizes the advance adjustment of refrigeration load and the coordinated distribution among cold storages, and enhances the adaptability of the control strategy to the change of the equipment health state.

Description

Technical Field

[0001] This invention relates to the field of energy management and control of cold storage clusters, and in particular to a collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning. Background Technology

[0002] In cold storage clusters consisting of multiple cold storage units, each unit typically obtains its cooling capacity by sharing compressor units, refrigeration piping, and power distribution. When the cooling demand of multiple cold storage units increases simultaneously within the same time period, the operating load of the shared compressor units rises rapidly, and the total power consumption of the cold storage cluster also increases accordingly, easily forming significant demand peaks. Existing cold storage control systems adjust compressors, dampers, and defrost based on cold storage temperature, refrigeration equipment operating status, refrigeration pressure, and equipment operating limits. While they can adjust the cooling capacity of individual cold storage units according to their current operating status, their ability to coordinate refrigeration resources between multiple cold storage units is limited, making it difficult to prevent multiple cold storage units from experiencing a surge in demand. The peak total power of the cold storage group caused by the concentrated increase of refrigeration load during similar periods; as an important part of cold storage refrigeration, the bearing grease of the evaporator fan is affected by speed, motor current, mechanical load, temperature and lubrication conditions during long-term operation, and will gradually oxidize, volatilize and thicken, entering a hidden deterioration state. This deterioration state is difficult to measure. Although the existing bearing vibration monitoring methods can extract the characteristic frequency and harmonic characteristics of the bearing, the vibration characteristic amplitude is affected by changes in operating conditions and mechanical defects of the bearing. It is easy to misjudge the enhanced vibration of the normal operating condition fluctuation as grease deterioration, reducing the accuracy of grease deterioration state identification.

[0003] Current technologies for energy-saving control of cold storage clusters only focus on temperature deviations, equipment operating boundaries, and current energy consumption data. They lack a means to integrate the health status of equipment and future degradation trends with the shared energy status of the cold storage cluster into control decisions. When individual cold storage units are controlled independently for energy saving, the lack of coordination between them may lead to increased refrigeration loads during the same period, resulting in superimposed loads on shared refrigeration equipment and an increase in the peak total power of the cold storage cluster. The deterioration status of bearing grease is a gradual and hidden degradation variable. Even if the current degree of deterioration can be identified, it is difficult to adjust the refrigeration load of cold storage units in advance and to transfer loads across cold storage units without combining historical degradation trajectories to predict future trends. Therefore, it is necessary to establish a control method that can identify the deterioration status of grease based on evaporator fan bearing vibration data, predict the deterioration trend, and integrate the health status of equipment, predicted health status, and shared refrigeration status of the cold storage cluster into multi-cold storage collaborative decision-making to achieve energy-saving operation of the cold storage cluster under demand constraints. Summary of the Invention

[0004] To overcome the difficulty of coordinating energy conservation across multiple cold storage facilities, the technical challenge is to provide a collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning.

[0005] The technical implementation scheme of the present invention is as follows: a collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning, comprising:

[0006] The cold storage group data acquisition and demand calculation module is used to collect data such as bearing vibration acceleration, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, and total power of the cold storage group, and calculate the average power of the demand sliding window.

[0007] The vibration characteristics and health benchmark calculation module is used to perform high-frequency bandpass filtering, envelope demodulation, and envelope spectrum calculation on bearing vibration acceleration data. It dynamically calculates the characteristic frequency of the target bearing based on the evaporator fan speed and extracts the characteristic frequency amplitude of the target bearing. It obtains the representative characteristic amplitude and representative characteristic energy index by calculating the root mean square average. It calculates the healthy vibration reference benchmark based on the healthy operation sample and the current operating conditions, and calculates the normalized characteristic energy ratio and mechanical defect suppression weight.

[0008] The grease degradation assessment and prediction module is used to calculate the degradation contribution and comprehensive degradation amount based on the normalized characteristic energy ratio of the working condition and the mechanical defect suppression weight, construct the bearing grease degradation degree index and perform validity processing; and obtain the predicted bearing grease degradation degree index by similarity matching between the historical trajectory of the effective bearing grease degradation degree index and the historical bearing grease degradation trajectory.

[0009] The multi-agent collaborative decision-making and safety constraint module is used to form the cold storage agent status by combining the effective bearing grease deterioration index, the predicted bearing grease deterioration index, the cold storage operation status, the evaporator fan operation status, the shared refrigeration status, and the cold storage group power status. Each cold storage agent outputs continuous control actions and defrosting control actions based on the cold storage agent status to coordinate shared refrigeration resources and enforce action safety constraints.

[0010] The control execution and status update module is used to execute control actions after the action safety constraints have been met, update the cold storage operation status and the energy status of the cold storage group, and form the input for the next control cycle.

[0011] Preferably, the step of collecting bearing vibration acceleration data, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, and total power of the cold storage group, and calculating the average power of the demand sliding window, includes: using vibration acceleration sensors, speed sensors or frequency converter feedback, current sensors, cold storage temperature sensors, cold storage controllers, damper position feedback, and shared compressor units installed on the evaporator fan bearing seats. The suction-side pressure sensor, the exhaust-side pressure sensor of the shared compressor unit, and the main incoming power metering device of the cold storage group collect data on bearing vibration acceleration, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature and defrosting status, cold storage damper opening, suction pressure of the shared compressor unit, condensing pressure of the shared compressor unit, and total power of the cold storage group, respectively. The average power of the demand sliding window is calculated based on the total power of the cold storage group, which is the average value of the total power of the cold storage group within the preset demand window length. All collected data are synchronized according to a unified timestamp.

[0012] Preferably, the step of performing high-frequency bandpass filtering, envelope demodulation, and envelope spectrum calculation on the bearing vibration acceleration data, dynamically calculating the target bearing characteristic frequency based on the evaporator fan speed, extracting the target bearing characteristic frequency amplitude, and obtaining representative characteristic amplitude and representative characteristic energy index through root mean square averaging includes: continuously sampling the bearing vibration acceleration data and dividing it into segments according to a preset vibration sampling frame length. A series of vibration sampling frames, The frequency band of the high-frequency impact signal is determined based on the evaporator fan bearing structure, the frequency response range of the vibration acceleration sensor, and the sampling frequency. Bandpass filtering is performed on each vibration sampling frame to obtain the high-frequency impact signal. An analytical signal is calculated from the high-frequency impact signal, and the modulus of the analytical signal is used as the envelope signal. A discrete Fourier transform is performed on the envelope signal to obtain the envelope spectrum. The characteristic frequency of the target bearing is dynamically calculated based on the evaporator fan speed and bearing structural parameters, including the number of rolling elements, rolling element diameter, pitch circle diameter, and contact angle. The amplitude of the characteristic frequency of the target bearing is extracted from the envelope spectrum. Within a preset sliding time window, the characteristic frequency of the same target bearing is... vibration sampling frame amplitude Perform the root mean square average according to the formula. Calculate the representative characteristic amplitude; obtain the representative characteristic energy index based on the square of the representative characteristic amplitude. .

[0013] Preferably, the step of calculating the healthy vibration reference benchmark based on healthy operating samples and current operating conditions includes: pre-establishing a healthy operating sample database to store operating data when the bearing grease is in a normal state and there are no obvious mechanical defects; the mechanical defects include damage to the bearing outer ring, damage to the bearing inner ring, and damage to the rolling elements; selecting from the healthy operating sample database the evaporator fan speed, evaporator fan motor operating current, difference between the cold storage temperature and the cold storage set temperature, cold storage damper opening, shared compressor unit suction pressure, and shared compressor unit condensing pressure as corresponding reference values ​​under healthy operating conditions; dividing the six current operating parameters by their corresponding reference values ​​to obtain six dimensionless variables for operating conditions, which are uniformly denoted as... Representative feature amplitudes and corresponding dimensionless variables of operating conditions are extracted from the healthy operation sample database to form healthy operation samples. Based on the healthy operation samples, the relationship between the dimensionless variables of centralized operating conditions and the representative feature amplitudes is fitted, and centralized operating condition variables are defined. To obtain a healthy vibration reference standard The formula is: ,in, When the dimensionless variables of the six operating conditions are all taken as their respective health reference values, the first... The healthy reference amplitude corresponding to the characteristic frequency of each target bearing For the first The characteristic frequency of the target bearing corresponds to the first... Parameters of a linear term model with dimensionless variables under centralized operating conditions. For the first The characteristic frequency of the target bearing corresponds to the first... The and the first Model parameters for interaction terms between dimensionless variables in a centralized operating condition. The first of the six dimensionless variables for operating conditions One variable, For the corresponding centralized operating conditions, dimensionless variables are used. The first of the six dimensionless variables for operating conditions One variable, For the corresponding centralized operating conditions, there are dimensionless variables.

[0014] Preferably, the calculation of the normalized characteristic energy ratio and mechanical defect suppression weight includes: converting the representative characteristic amplitude... Compared with healthy vibration reference standards Normalization is performed to obtain the normalized characteristic energy ratio under the operating condition. The formula is: ,in The value is a preset, extremely small positive number with dimensions consistent with the representative characteristic amplitude; the concentration of mechanical defect features is calculated based on the representative characteristic energy index. The concentration degree of mechanical defect features is equal to the ratio of the sum of the representative characteristic energy indices corresponding to the mechanical defect feature frequencies to the sum of the representative characteristic energy indices corresponding to the feature frequencies of all target bearings. The set of mechanical defect feature frequencies is composed of the target bearing feature frequencies corresponding to the mechanical defects. Mechanical defect suppression weights are calculated based on the concentration degree of mechanical defect features. The formula is: ,in This is the preset mechanical defect suppression coefficient; when the concentration of mechanical defect features is continuous When a control cycle is greater than or equal to a preset mechanical defect judgment threshold, the evaporator fan is determined to be in an abnormal mechanical defect state. It is a preset positive integer.

[0015] Preferably, the step of calculating the degradation contribution and overall degradation amount based on the normalized characteristic energy ratio of operating conditions and the mechanical defect suppression weight, constructing a bearing grease degradation index, and performing effectiveness processing includes: calculating the degradation contribution of each target bearing characteristic frequency based on the normalized characteristic energy ratio of operating conditions and the mechanical defect suppression weight. The formula is: ,in For the first Nonlinear sensitivity index of the characteristic frequency of a target bearing ; Calculate the overall degradation amount The formula is: ;in , The first The and the first The characteristic frequency weights of each target bearing are all non-negative and sum to 1. These are the nonlinear interaction coefficients between features. For the first The degradation contribution of each target bearing characteristic frequency; candidate values ​​for the bearing grease degradation index are calculated based on the overall degradation amount. The formula is: ,in The preset global nonlinear index is used; when the evaporator fan is not in a mechanical defect abnormal state, the candidate value of the bearing grease deterioration index is determined as the effective bearing grease deterioration index; when the evaporator fan is in a mechanical defect abnormal state, the effective bearing grease deterioration index of the previous moment and the candidate value of the current bearing grease deterioration index are calculated according to the formula. A smooth update was performed to obtain an effective bearing grease degradation index; among which... For smoothing coefficients, This represents the index of the degree of deterioration of the effective bearing grease at the previous moment.

[0016] Preferably, the step of obtaining the predicted bearing grease degradation index by similarity matching between the historical trajectory of the effective bearing grease degradation index and the historical bearing grease degradation trajectory includes: continuously recording the degradation index of each evaporator fan bearing. The effective bearing grease degradation index at each time point was recorded simultaneously with the above. The dimensionless variables of the operating conditions corresponding to each time point are used to form the historical trajectory of the current bearing grease degradation index. A historical bearing grease degradation trajectory database is pre-established to store the historical trajectory of the bearing grease degradation index during the gradual degradation process, the corresponding historical trajectory of the operating conditions, and the bearing grease degradation index after a preset future time. H historical bearing grease degradation trajectories are obtained from the historical bearing grease degradation trajectory database, where H is a preset positive integer. Each historical bearing grease degradation trajectory contains the same variables, time length, and normalization method as the current bearing grease degradation index historical trajectory. The weighted Euclidean distance between the current bearing grease degradation index historical trajectory and each historical bearing grease degradation trajectory is calculated. Similarity is calculated based on weighted Euclidean distance. ,in The similarity scale parameter is set as a preset value; the bearing grease deterioration index corresponding to each historical trajectory after a preset future time is weighted and averaged according to similarity to obtain the predicted bearing grease deterioration index. The formula is: ,in For the preset future time step, For the current moment The subsequent predicted time, It is a preset minimum positive number.

[0017] Preferably, the step of forming the cold storage intelligent agent status from the effective bearing grease deterioration index, the predicted bearing grease deterioration index, the cold storage operating status, the evaporator fan operating status, the shared refrigeration status, and the cold storage group power status includes: defining each cold storage in the cold storage group as a cold storage intelligent agent; the first Local observations of the cold storage intelligent agent include cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, evaporator fan speed, evaporator fan motor operating current, effective bearing grease degradation index, and predicted bearing grease degradation index; based on the natural logarithm of the ratio of the shared compressor unit's condensing pressure to the shared compressor unit's suction pressure, according to the formula... Calculate the shared refrigeration pressure state, where To share the condensing pressure of the compressor unit, To share the suction pressure of the compressor unit; based on the ratio of the total electrical power of the cold storage group to the reference value of the cold storage group's power, according to the formula... Calculate the dimensionless power state of the cold storage group, where This is a reference value for the power of cold storage groups. The total electrical power of the cold storage group is represented by the observations of each cold storage agent. The observations of the cold storage agents are combined with the local observations of each cold storage agent and the shared refrigeration pressure state, the dimensionless power state of the cold storage group, the average power of the demand sliding window, and the demand over-limit to form the observations of the cold storage agents. During the centralized training phase, the observations of each cold storage agent are combined into the global state.

[0018] Preferably, each cold storage intelligent agent outputs continuous control actions and defrost control actions based on its status to coordinate shared refrigeration resources and enforce action safety constraints. This includes: each cold storage intelligent agent outputting the cold storage set temperature adjustment, cold storage refrigeration demand adjustment, and cold storage damper opening adjustment as continuous control actions, and outputting a defrost start request as a defrost control action; when multiple cold storages share a shared compressor unit, the cold storage refrigeration demand adjustment output by each cold storage intelligent agent is input into the shared refrigeration resource coordination mechanism. The shared refrigeration resource coordination mechanism is based on the refrigeration demand adjustment of each cold storage, the suction pressure of the shared compressor unit, the condensing pressure of the shared compressor unit, and the shared compressor unit's refrigeration demand adjustment. The cold storage capacity, the safety margin of each cold storage temperature, and the power status of the cold storage group determine the final operating command of the shared compressor unit; action safety constraints are implemented for continuous control actions and defrosting control actions, including the upper and lower limits of the operating frequency of the cold storage compressor, the upper and lower limits of the cold storage damper opening, the upper and lower limits of the cold storage set temperature, the upper and lower limits of the predicted cold storage temperature, the pressure boundary of the shared compressor unit, and the boundary of the shared refrigeration capacity; when the predicted bearing grease deterioration index is greater than or equal to the warning threshold of the bearing grease deterioration index but less than the critical threshold of the bearing grease deterioration index, and the predicted cold storage temperature is close to the upper limit of the cold storage temperature, the action of reducing the refrigeration capacity is restricted.

[0019] Preferably, the execution of the control action after safety constraints, updating the cold storage operating status and the energy status of the cold storage group, and forming the input for the next control cycle includes: sending the control action after safety constraints to the cold storage actuator; generating new cold storage temperature, cold storage set temperature, cold storage refrigeration equipment operating status, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, total power of the cold storage group, and new bearing vibration acceleration data after the control action is executed; calculating the reinforcement learning reward based on the new total power of the cold storage group, the average power of the demand sliding window, the cold storage temperature, the predicted bearing grease degradation index, and the control action amplitude, and the reinforcement learning reward is calculated according to the formula. Calculations, including the requirement exceeding the limit item. , , This represents the dimensionless power state of the cold storage group. , , Non-negative reward weights To predictively adjust the reward, new operating data and vibration acceleration data are used as inputs for the next control cycle.

[0020] Beneficial effects: This invention dynamically calculates the bearing characteristic frequency based on the real-time rotational speed of the evaporator fan and extracts the amplitude of the target bearing characteristic frequency, reducing frequency drift errors caused by fixed frequency position extraction and ensuring stable identification of grease degradation under variable speed conditions. By establishing a nonlinear healthy vibration reference benchmark that varies with operating conditions using evaporator fan speed, motor current, the difference between cold storage temperature and set temperature, damper opening, and the suction and discharge pressure of the shared compressor unit, representative characteristic amplitudes are normalized to the reference benchmark, reducing interference from normal operating condition fluctuations on the degradation degree assessment. Furthermore, by calculating the concentration of mechanical defect features and forming an exponential suppression weight, the local feature enhancement caused by damage to the bearing outer ring, inner ring, and rolling elements is suppressed, reducing grease degradation caused by mechanical defects. To mitigate misjudgments, a bounded and continuous bearing grease degradation index is constructed by converting the normalized energy ratio of the characteristic frequencies of multiple target bearings into a nonlinear degradation contribution and superimposing the interaction terms between features. This allows for continuous differentiation between normal, warning, and critical states. Furthermore, by performing nonlinear similarity matching between the historical trajectory of the current bearing grease degradation index and the historical degradation trajectory, a predicted value is obtained after a preset time, enabling the control strategy to adjust the refrigeration load in advance based on the degradation trend. Finally, by combining the effective bearing grease degradation index, predicted value, cold storage operating status, shared refrigeration status, and cold storage group power status into a cold storage intelligent agent observation, and employing shared refrigeration resource coordination and action safety constraints, multiple cold storage facilities can achieve coordinated energy-saving operation while meeting temperature, pressure, and demand constraints. Attached Figure Description

[0021] Figure 1 This is a structural diagram of the collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning, as described in this invention.

[0022] Figure 2 This is a flowchart of the vibration signal feature extraction method of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: A collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning, such as... Figure 1 and Figure 2 As shown, it includes:

[0025] The cold storage group data acquisition and demand calculation module is used to collect data such as bearing vibration acceleration, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, and total power of the cold storage group, and calculate the average power of the demand sliding window.

[0026] The vibration characteristics and health benchmark calculation module is used to perform high-frequency bandpass filtering, envelope demodulation, and envelope spectrum calculation on bearing vibration acceleration data. It dynamically calculates the characteristic frequency of the target bearing based on the evaporator fan speed and extracts the characteristic frequency amplitude of the target bearing. It obtains the representative characteristic amplitude and representative characteristic energy index by calculating the root mean square average. It calculates the healthy vibration reference benchmark based on the healthy operation sample and the current operating conditions, and calculates the normalized characteristic energy ratio and mechanical defect suppression weight.

[0027] The grease degradation assessment and prediction module is used to calculate the degradation contribution and comprehensive degradation amount based on the normalized characteristic energy ratio of the working condition and the mechanical defect suppression weight, construct the bearing grease degradation degree index and perform validity processing; and obtain the predicted bearing grease degradation degree index by similarity matching between the historical trajectory of the effective bearing grease degradation degree index and the historical bearing grease degradation trajectory.

[0028] The multi-agent collaborative decision-making and safety constraint module is used to form the cold storage agent status by combining the effective bearing grease deterioration index, the predicted bearing grease deterioration index, the cold storage operation status, the evaporator fan operation status, the shared refrigeration status, and the cold storage group power status. Each cold storage agent outputs continuous control actions and defrosting control actions based on the cold storage agent status to coordinate shared refrigeration resources and enforce action safety constraints.

[0029] The control execution and status update module is used to execute control actions after they have been constrained by safety requirements, update the operating status of the cold storage and the energy status of the cold storage group, and form the input for the next control cycle.

[0030] By using vibration acceleration sensors, speed sensors or frequency converter feedback, current sensors, cold storage temperature sensors, cold storage controllers, damper position feedback, shared compressor unit suction side pressure sensors, shared compressor unit exhaust side pressure sensors, and a cold storage group's total incoming power metering device installed on the evaporator fan bearing housing, the following data are collected: bearing vibration acceleration data, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature and defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, and total power of the cold storage group. The average power of the demand sliding window is calculated based on the total power of the cold storage group, which is the average value of the total power of the cold storage group within a preset demand window length. All collected data are synchronized using a unified timestamp.

[0031] It should be explained that the bearing vibration acceleration data is collected by piezoelectric vibration acceleration sensors installed on the bearing housing of each evaporator fan, and the data is converted from analog to digital to form the original time series. The unit is m / s²; the evaporator fan speed is obtained from the feedback signal of the encoder, speed sensor, or evaporator fan frequency converter, and is denoted as . The unit is r / min; the operating current of the evaporator fan motor is measured by a current sensor connected in series in the motor power supply circuit, and is denoted as . The unit is A; the cold storage temperature is collected by a temperature sensor installed inside the cold storage, and is denoted as . The unit is °C; the set temperature and defrost status of the cold storage are provided by the cold storage controller, wherein the set temperature is denoted as . Defrosting status A value of 1 indicates defrosting operation, and a value of 0 indicates no defrosting; the opening degree of the cold storage damper is obtained from the position feedback of the damper actuator, and is denoted as... The unit is %; the suction pressure and condensing pressure of the shared compressor unit are measured by pressure sensors installed on the suction manifold and exhaust manifold, respectively, and are denoted as %. and The unit is kPa and absolute pressure is used; the total electrical power of the cold storage group is measured by the main incoming power metering device of the cold storage group, and is denoted as... The unit is kW; for the operating frequency of the cold storage compressor, the independent compressor cold storage obtains the frequency from the frequency converter. When calculating the average power of the demand sliding window, first set the preset demand window length according to the power supply demand billing cycle. The window length is greater than one control cycle, and the window contains Each total power sampling point, according to the formula Calculation, where The sampling time within the window; the preset required window length The demand calculation period is determined according to the power supply contract for the cold storage group or the electricity metering device, and the actual length of the demand calculation period is taken as... When the electricity demand billing cycle is a fixed time length, the fixed time length is used, and the setting is made so that... The value is an integer multiple of the control cycle closest to the control cycle. The sliding window is updated once every control cycle, and the average power of the earliest control cycle in the window is replaced by the average power of the new control cycle. The time synchronization of all collected data is achieved by timestamping with a unified clock source. The sampling frequency difference between high-frequency vibration data and low-frequency operation data is handled by aligning the timestamp of the low-frequency operation data to the start time of the vibration sampling frame. The time scale difference between the control cycle and the vibration sampling cycle is taken as the latest operation data in the corresponding time window within a control cycle as the input of the cold storage agent.

[0032] The bearing vibration acceleration data is continuously sampled and divided into segments according to a preset vibration sampling frame length. A series of vibration sampling frames, The frequency band of the high-frequency impact signal is determined based on the evaporator fan bearing structure, the frequency response range of the vibration acceleration sensor, and the sampling frequency. Bandpass filtering is performed on each vibration sampling frame to obtain the high-frequency impact signal. An analytical signal is calculated from the high-frequency impact signal, and the modulus of the analytical signal is used as the envelope signal. A discrete Fourier transform is performed on the envelope signal to obtain the envelope spectrum. The characteristic frequency of the target bearing is dynamically calculated based on the evaporator fan speed and bearing structural parameters, including the number of rolling elements, rolling element diameter, pitch circle diameter, and contact angle. The amplitude of the characteristic frequency of the target bearing is extracted from the envelope spectrum. Within a preset sliding time window, the characteristic frequency of the same target bearing is... vibration sampling frame amplitude Perform the root mean square average according to the formula. Calculate the representative characteristic amplitude; obtain the representative characteristic energy index based on the square of the representative characteristic amplitude. .

[0033] It should be explained that the vibration signal processing begins with continuous sampling of the bearing vibration acceleration data at a high-frequency sampling rate. The sampling frequency is determined by the upper limit of the frequency response of the vibration accelerometer and the Nyquist sampling theorem. The sampling sequence is then segmented according to a preset vibration sampling frame length. Each frame contains a fixed number of sampling points while maintaining the continuity of adjacent frames on the time axis, thereby obtaining... A series of vibration sampling frames, in which The preset positive integer value is determined based on the matching relationship between the time required to complete all subsequent envelope spectrum calculations within a single control cycle and the hardware computing power. The process of determining the high-frequency impact band is as follows: First, the range of structural resonance frequencies caused by bearing defect impacts is estimated based on the number, diameter, and pitch circle diameter of the evaporator fan bearing. Then, combined with the nominal resonant frequency of the vibration acceleration sensor and the upper limit of the effective frequency band that can be retained from the actual sampling frequency, the frequency band dominated by low-frequency rotational noise and electromagnetic interference is eliminated, and the remaining frequency band is taken as the high-frequency impact band. When performing bandpass filtering on each vibration sampling frame, a finite impulse response filter is used to suppress frequency components outside the high-frequency impact band. The passband cutoff frequency and stopband cutoff frequency of the filter are set according to the upper and lower boundaries of the high-frequency impact band, respectively, to obtain the high-frequency impact signal. The envelope signal is obtained through Hilbert transform for high-frequency impulse signals. The analytic signal is calculated, where the real part is the high-frequency impulse signal itself, and the imaginary part is the result of the high-frequency impulse signal after Hilbert transform. The envelope signal is obtained by taking the square root of the sum of the squares of the real and imaginary parts of the analytic signal. The envelope signal removes the high-frequency carrier while retaining the low-frequency envelope characteristics caused by the impact; subsequently, a Discrete Fourier Transform is performed on the envelope signal. The input of the Discrete Fourier Transform is all the sampling points of the envelope signal within one vibration sampling frame, and the output is the frequency domain envelope spectrum. The frequency resolution of the envelope spectrum is determined by the time-domain duration and sampling frequency corresponding to the vibration sampling frame length; the calculation of the characteristic frequency of the target bearing is based on the evaporator fan speed, first by setting the evaporator fan speed... Divide by sixty to get the frequency. The unit is Hz, and then based on the number of rolling elements in the bearing structural parameters. Rolling element diameter d, pitch circle diameter and contact angle The outer ring damage characteristic frequency BPFO, inner ring damage characteristic frequency BPFI, rolling element damage characteristic frequency BSF, and cage characteristic frequency FTF were calculated separately. The expressions for each characteristic frequency are expressed in terms of rotational frequency. Multiply by the dimensionless coefficients composed of the above bearing structural parameters to obtain the results, respectively according to... , , and Calculation, where For the evaporator fan frequency, For the number of rolling elements, The diameter of the rolling element, The diameter of the pitch circle. The contact angle is used; when the evaporator fan speed changes, the target bearing characteristic frequency is dynamically updated over time; when extracting the target bearing characteristic frequency amplitude from the envelope spectrum, a frequency search window is set with the target bearing characteristic frequency as the center, and the half-width of the search window is taken as an integer multiple of the envelope spectrum frequency resolution. The maximum value of the envelope spectrum amplitude within the search window is taken as the target bearing characteristic frequency amplitude. Within a preset sliding time window, the characteristic frequency corresponding to the same target bearing... The amplitude of each vibration sampling frame is averaged using the root mean square method. ,in Indicates the representative characteristic amplitude. The number of vibration sampling frames within a preset sliding time window. Indicates the first The characteristic frequency of the target bearing is at the first The amplitude of each vibration sampling frame is squared to ensure a greater contribution of the amplitude during averaging, and the root mean square (RMS) operation is used to obtain statistical results consistent with the original amplitude dimensions. The length of the preset sliding time window is consistent with the control cycle of the cold storage agent, and the window slides forward once after each vibration sampling frame, so that the representative characteristic amplitude can reflect the stability level of the target bearing's characteristic frequency amplitude in the most recent period. Finally, the representative characteristic energy index is obtained based on the square of the representative characteristic amplitude. ,in It has the dimension of energy and is used in subsequent steps to measure the concentration of vibration energy corresponding to the characteristic frequency of each target bearing in the total vibration energy.

[0034] A pre-established healthy operation sample database is used to store operating data when the bearing grease is in a normal state and there are no obvious mechanical defects. These mechanical defects include damage to the bearing outer ring, inner ring, and rolling elements. From the healthy operation sample database, the evaporator fan speed, evaporator fan motor operating current, the difference between the cold storage temperature and the set cold storage temperature, the cold storage damper opening, the shared compressor unit suction pressure, and the shared compressor unit condensing pressure are selected as corresponding reference values. The six current operating parameters are divided by their respective reference values ​​to obtain six dimensionless variables for operating conditions, which are uniformly denoted as... Representative feature amplitudes and corresponding dimensionless variables of operating conditions are extracted from the healthy operation sample database to form healthy operation samples. Based on the healthy operation samples, the relationship between the dimensionless variables of centralized operating conditions and the representative feature amplitudes is fitted, and centralized operating condition variables are defined. To obtain a healthy vibration reference standard The formula is: ,in, When the dimensionless variables of the six operating conditions are all taken as their respective health reference values, the first... The healthy reference amplitude corresponding to the characteristic frequency of each target bearing For the first The characteristic frequency of the target bearing corresponds to the first... Parameters of a linear term model with dimensionless variables under centralized operating conditions. For the first The characteristic frequency of the target bearing corresponds to the first... The and the first Model parameters for interaction terms between dimensionless variables in a centralized operating condition. The first of the six dimensionless variables for operating conditions One variable, For the corresponding centralized operating conditions, dimensionless variables are used. The first of the six dimensionless variables for operating conditions One variable, For the corresponding centralized operating conditions, there are dimensionless variables.

[0035] It should be explained that the healthy operation sample database was built offline. The data source is the screening and archiving of historical operation records of the same model of evaporator fan after the bearing grease was replaced, and vibration testing confirmed that there was no abnormal increase in the characteristic frequency of bearing outer ring damage, bearing inner ring damage, and rolling element damage. The bearing grease being in a normal state means that the cumulative operating time of the evaporator fan since the most recent bearing grease replacement or replenishment has not exceeded the preset normal lubrication operation time. The equipment maintenance records show no obvious grease leakage, severe grease deficiency, or abnormal grease replenishment, and the corresponding operating data does not show any confirmed bearing mechanical defects; the normal lubrication operation time... The normal service life of the bearing grease is determined according to the manufacturer's specifications for the same model of evaporator fan. If the manufacturer does not provide such a normal service life, it is determined based on the median of the operating time from the moment of grease replacement or replenishment to the first confirmation that the grease has entered a warning state in historical operating records. The operating period meeting these conditions is defined as the normal operating period for the bearing grease. Each record includes the evaporator fan speed, evaporator fan motor operating current, the difference between the cold storage temperature and the set cold storage temperature, the cold storage damper opening, the shared compressor unit suction pressure, the shared compressor unit condensing pressure, and representative characteristics corresponding to the characteristic frequencies of each target bearing at the corresponding sampling time. The amplitude, along with the bearing grease filling time, cumulative operating time, and equipment maintenance records, serves as an auxiliary basis for judging the health status. Data sets of six types of operating parameters under healthy operating conditions are extracted from the healthy operation sample database. For evaporator fan speed, evaporator fan motor operating current, cold storage damper opening, shared compressor unit suction pressure, and shared compressor unit condensing pressure, the median of the positive values ​​in the corresponding healthy operation data is taken as the corresponding reference value. For the difference between the cold storage temperature and the set cold storage temperature, the absolute value of the temperature deviation in the healthy operation data is first taken, and then the median of the positive values ​​is taken as the temperature difference reference value. Six corresponding reference values ​​were obtained. , , , , and Each reference value is positive. Data sets of the above six types of operating parameters under healthy operating conditions are extracted from the healthy operating sample database. For each set, the truncated mean method is used to remove outliers at both ends, and the mean is taken to obtain six corresponding reference values. , , , , and And all reference values ​​remain positive to ensure that subsequent dimensionless conversion is meaningful; the evaporator fan speed obtained in the current control cycle Evaporator fan motor operating current The difference between the cold storage temperature and the set temperature of the cold storage Cold storage air door opening Shared compressor unit suction pressure and shared compressor unit condensing pressure Dividing each of these by the corresponding reference values ​​yields six dimensionless variables for different operating conditions. to These correspond to speed, current, temperature difference, damper opening, suction pressure, and condensing pressure, respectively; the temperature difference variable is calculated by dividing the temperature difference itself by... , , , , and The ratio of the current measured value to its respective reference value is uniformly denoted as . , The range is from 1 to 6; the healthy operation samples are formed by extracting representative feature amplitudes from each record in the healthy operation sample database. and the corresponding six dimensionless variables This constitutes a set of data pairs used for model fitting; before fitting, the natural logarithm of the representative feature amplitudes is taken, and the six dimensionless variables of the operating conditions are... Convert to dimensionless variables of centralized operating conditions Transforming the benchmark model in exponential form into one about The linear-plus-interaction term form aims to minimize the mean squared error between the logarithm of the measured representative feature amplitudes in the healthy sample and the logarithm of the model prediction. A multiple linear regression algorithm with a quadratic interaction term is used to solve for the model parameters. The model intercept directly corresponds to the logarithmic health reference amplitude when all six operating condition variables are at their respective health reference values, thus obtaining the... Standard healthy operating condition reference amplitude of the characteristic frequency of the target bearing Linear term model parameters and interaction item model parameters Before fitting, the natural logarithm of the representative feature amplitudes is taken, transforming the exponential form of the baseline model into a model about... The linear-plus-interaction term form aims to minimize the mean squared error between the logarithm of the measured representative feature amplitudes in the healthy sample and the logarithm of the model prediction. A multiple linear regression algorithm with a quadratic interaction term is used to solve for the model parameters, yielding the th... Healthy reference amplitude of the characteristic frequency of each target bearing And linear term model parameters and interaction item model parameters Healthy Vibration Reference Standard The calculation formula is ,in Indicates the first The reference amplitude of the characteristic frequency of each target bearing under standard healthy operating conditions, the exponential function ensures that the healthy vibration reference is always positive; Describe the independent linear effects of dimensionless variables under six operating conditions on the amplitude of representative characteristics. Describe the nonlinear correction effect of pairwise coupling between variables under different operating conditions on the amplitude of representative characteristics. Represents the dimensionless variable of the six operating conditions. 1 variable, and and Iterate through all pairwise combinations; the resulting healthy vibration reference benchmark can be dynamically adjusted according to the current operating conditions, and the vibration amplitude fluctuations caused by changes in normal operating conditions will be deducted from the degradation assessment during subsequent normalization.

[0036] Representative characteristic amplitude Compared with healthy vibration reference standards Normalization is performed to obtain the normalized characteristic energy ratio under the operating condition. The formula is: ,in The value is a preset, extremely small positive number with dimensions consistent with the representative characteristic amplitude; the concentration of mechanical defect features is calculated based on the representative characteristic energy index. The concentration degree of mechanical defect features is equal to the ratio of the sum of the representative characteristic energy indices corresponding to the mechanical defect feature frequencies to the sum of the representative characteristic energy indices corresponding to the feature frequencies of all target bearings. The set of mechanical defect feature frequencies is composed of the target bearing feature frequencies corresponding to the mechanical defects. Mechanical defect suppression weights are calculated based on the concentration degree of mechanical defect features. The formula is: ,in This is the preset mechanical defect suppression coefficient; when the concentration of mechanical defect features is continuous When a control cycle is greater than or equal to a preset mechanical defect judgment threshold, the evaporator fan is determined to be in an abnormal mechanical defect state. It is a preset positive integer.

[0037] It needs to be explained that the normalized characteristic energy ratio of the operating condition Used to compare measured vibration energy with a health reference level under current operating conditions, the calculation formula is as follows: , of which molecules The denominator is the representative characteristic amplitude obtained by root mean square averaging, reflecting the current actual vibration intensity; Take the healthy vibration reference standard and the preset minimum positive number The larger value in This is used to ensure that the denominator is not zero and that the dimensions are consistent with the representative characteristic amplitude; The determination method is as follows: extract all representative feature amplitudes from the healthy operation sample database, count the smallest non-zero value, and multiply the smallest non-zero value by a preset scaling factor. The scaling factor is determined by data precision and the need to prevent division by zero. The squaring operation converts the amplitude ratio into an energy ratio, making... In a healthy state, it fluctuates around a value of 1. When the actual vibrational energy exceeds the healthy reference level... Greater than 1, when below the health reference level Less than 1; Concentration of mechanical defect features The calculation process is as follows: first, define the set of characteristic frequencies of mechanical defects. The set consists of the target bearing characteristic frequencies and preset order harmonics corresponding to bearing outer ring damage, bearing inner ring damage, and rolling element damage, representing the complete set of target bearing characteristic frequencies. Includes cage characteristic frequencies and characteristic frequencies used in the evaluation. for A subset; the numerator is Representative eigenenergy index corresponding to all characteristic frequencies The sum of the denominators is Representative eigenenergy index corresponding to all characteristic frequencies The sum and ratio reflect the proportion of defect-related frequency energy in the total vibration energy; when the denominator is zero or less than a preset minimum energy threshold, the concentration of mechanical defect features is set to zero; mechanical defect suppression weight. The calculation formula is ,in The preset mechanical defect suppression coefficient, The determination method is as follows: using samples from the mechanical defect sample database that have been confirmed to have bearing outer ring damage, inner ring damage, or rolling element damage but with normal grease condition, and samples with degraded grease but no mechanical defects, the values ​​of the two types of samples are calculated. The distribution differences are determined through a one-dimensional search, with the goal of minimizing the misjudgment rate of grease degradation caused by mechanical defects. The value of ; When greater than zero, The bigger The closer the value is to zero, the more the contribution of the defect frequency is suppressed in subsequent degradation contribution calculations; the determination of abnormal mechanical defect states adopts a continuous confirmation mechanism, and when the concentration of mechanical defect features is high... continuous The number of control cycles is greater than or equal to the preset mechanical defect judgment threshold. At that time, it was determined that the evaporator fan was in an abnormal state of mechanical defect; The determination method is based on the fluctuation duration of the concentration of defect features in the mechanical defect sample and the requirements for misjudgment risk, and it is necessary to ensure that short-term disturbances will not trigger abnormal states. The method for determining this is to statistically analyze the healthy operation sample database. The distribution is used, and the upper quantile of the distribution is taken as the threshold. The quantile level is set according to the trade-off between false alarm rate and false negative rate. Continuous judgment is implemented by a counter. In each control cycle, the counter is incremented by one when the threshold condition is met and reset to zero when the threshold condition is not met. When the counter reaches a certain value... An abnormal state is triggered at that time.

[0038] The degradation contribution of each target bearing's characteristic frequency is calculated based on the normalized characteristic energy ratio under operating conditions and the mechanical defect suppression weight. The formula is: ,in For the first Nonlinear sensitivity index of the characteristic frequency of a target bearing ; Calculate the overall degradation amount The formula is: ;in , The first The and the first The characteristic frequency weights of each target bearing are all non-negative and sum to 1. These are the nonlinear interaction coefficients between features. For the first The degradation contribution of each target bearing characteristic frequency; candidate values ​​for the bearing grease degradation index are calculated based on the overall degradation amount. The formula is: ,in The preset global nonlinear index is used; when the evaporator fan is not in a mechanical defect abnormal state, the candidate value of the bearing grease deterioration index is determined as the effective bearing grease deterioration index; when the evaporator fan is in a mechanical defect abnormal state, the effective bearing grease deterioration index of the previous moment and the candidate value of the current bearing grease deterioration index are calculated according to the formula. A smooth update was performed to obtain an effective bearing grease degradation index; among which... For smoothing coefficients, This represents the index of the degree of deterioration of the effective bearing grease at the previous moment.

[0039] It should be explained that the aforementioned degradation contribution The calculation combines the normalized characteristic energy ratio of the operating condition and the mechanical defect suppression weight to quantify the independent contribution of the characteristic frequency of a single target bearing to the deterioration of the bearing grease under the current condition. The calculation formula is as follows: ,in The mechanical defect suppression weight obtained in the previous stage is reduced when the concentration of mechanical defect features increases, so that the enhancement of feature frequency caused by mechanical defects will not be misjudged as grease deterioration. In For the first The nonlinear sensitivity index of the characteristic frequency of each target bearing was obtained by using field tracking data to statistically analyze the change curve of the characteristic frequency energy ratio of each target bearing with the degree of grease deterioration. The fitted value of the slope of the curve on logarithmic coordinates was used as the... , Greater than zero; This is used to convert the nonlinear growth of the energy ratio into a contribution on a logarithmic scale, minus The threshold judgment is made when When less than or equal to 1, the logarithmic value does not exceed positive operation Set the degradation contribution to zero, and when When the value is greater than 1, the degradation contribution increases monotonically with the increase of energy ratio; the overall degradation amount It consists of two parts, and the calculation formula is: ,in The weighted linear summation of the degradation contributions of each target bearing's characteristic frequency reflects the cumulative effect of the degradation degree of multiple characteristic frequencies. This is a nonlinear interaction term between the characteristic frequencies of different target bearings, used to capture the synergistic degradation effect generated when multiple characteristic frequencies are enhanced simultaneously; and The first The and the first The weights of the target bearing characteristic frequencies are non-negative and sum to 1. The method of obtaining them is determined based on the stability and characterization ability of different target bearing characteristic frequencies in bearing grease degradation data. The ratio of the inter-class divergence to the intra-class divergence of each characteristic frequency between samples in normal and deteriorated states is calculated, and then the ratio is normalized as the weight. The nonlinear interaction coefficients between features are obtained by selecting bearing grease degradation samples where multiple target bearing characteristic frequencies simultaneously increase, in order to minimize the overall degradation amount. The error in identifying the deterioration stage of lubricating grease was determined through regression analysis; This is used to perform a geometric mean on the weights of two feature frequencies, so that the interaction is matched with their respective weights. For the first The degradation contribution of each target bearing characteristic frequency; candidate values ​​for the bearing grease degradation index. The formula is based on the overall degradation amount. ,in The preset global nonlinear index is obtained by determining the bearing grease degradation index distribution under normal, warning, and critical bearing grease conditions. The objective is to maximize the separability of candidate values ​​across the three state intervals, and this is achieved through a one-dimensional search. The value of ; because Non-negative, The value range is between 0 and 1, with candidate values ​​always falling within this range and tending towards 1 as the overall degradation increases. During the effectiveness processing stage, the final effective bearing grease degradation index is determined based on the mechanical defect anomaly state judgment results. When the evaporator fan is not in a mechanical defect anomaly state and the current vibration data is valid, the candidate value of the bearing grease degradation index is directly determined as the effective bearing grease degradation index. When the evaporator fan is in a mechanical defect anomaly state, a smooth update method is used to suppress instantaneous fluctuations caused by defect impacts. The calculation formula is... This is the index of the degree of deterioration of the bearing grease at the previous moment. For smoothing coefficients, The acquisition method involves selecting historical vibration data under abnormal mechanical defect conditions, aiming to minimize the mean square error between the smoothed exponent and the manually determined degree of grease deterioration, and determining the optimal method through a one-dimensional search. The value of , The closer the current candidate value is to 1, the smaller its impact on the effective value. When the current vibration data is invalid, the effective bearing grease deterioration index of the previous moment remains unchanged to ensure that the effective values ​​used for subsequent historical trajectory construction and prediction are continuous and stable.

[0040] Continuously record each evaporator fan bearing The effective bearing grease degradation index at each time point was recorded simultaneously with the above. The dimensionless variables of the operating conditions corresponding to each time point are used to form the historical trajectory of the current bearing grease degradation index. A historical bearing grease degradation trajectory database is pre-established to store the historical trajectory of the bearing grease degradation index during the gradual degradation process, the corresponding historical trajectory of the operating conditions, and the bearing grease degradation index after a preset future time. H historical bearing grease degradation trajectories are obtained from the historical bearing grease degradation trajectory database, where H is a preset positive integer. Each historical bearing grease degradation trajectory contains the same variables, time length, and normalization method as the current bearing grease degradation index historical trajectory. The weighted Euclidean distance between the current bearing grease degradation index historical trajectory and each historical bearing grease degradation trajectory is calculated. Similarity is calculated based on weighted Euclidean distance. ,in The similarity scale parameter is set as a preset value; the bearing grease deterioration index corresponding to each historical trajectory after a preset future time is weighted and averaged according to similarity to obtain the predicted bearing grease deterioration index. The formula is: ,in For the preset future time step, For the current moment The subsequent predicted time, It is a preset minimum positive number.

[0041] It should be noted that the current historical trajectory of the bearing grease deterioration index is constructed using a single evaporator fan bearing as the object, continuously recording the most recent data for that bearing. Effective bearing grease degradation index for each control cycle At the same time, the dimensionless variables of the six types of operating conditions corresponding to each control cycle are... to Stored together according to the same time index, forming a historical trajectory of the current bearing grease deterioration index, which is composed of the time series of the effective bearing grease deterioration index and the time series of dimensionless variables of operating conditions. ;in The preset positive integer is determined by statistically analyzing the distribution of the number of control cycles required for the bearing grease to progress from a normal state to a critical state, and then taking the minimum number of cycles required to cover the main deterioration trend changes from the distribution. This ensures that the historical trajectory contains sufficient degradation information without introducing excessive outdated data due to its length. The historical bearing grease degradation trajectory database is established offline, with data sourced from historical field tracking records. Each historical bearing grease degradation trajectory records the effective bearing grease degradation degree index sequence during the gradual degradation process of the bearing grease from a healthy state, the corresponding dimensionless variable sequence of operating conditions, and a preset future time step from the end of the sequence. The bearing grease degradation index obtained from the subsequent simulation Each trajectory in the database is arranged according to the same variable order and time length as the current historical trajectory. Organized using normalization methods, when the length of the original historical trajectory exceeds... Extract the portion aligned with the end of the current trajectory. At a certain point in time, when insufficient The aforementioned trajectory will not be included in subsequent matching; when retrieving historical bearing grease degradation trajectories from the historical bearing grease degradation trajectory database, the... The preset upper limit for the number of historical degradation trajectories is set. Initial screening is performed based on the Euclidean distance between the dimensionless variable of the current operating condition and the corresponding operating condition of each historical trajectory. The preset operating condition screening threshold is determined based on the distribution of Euclidean distances between historical trajectories. The upper quantile of the distance distribution is taken as the operating condition screening threshold. Historical trajectories with distances greater than the threshold are removed. Among the remaining historical trajectories, they are arranged in ascending order of Euclidean distance from the operating conditions, and the top-ranked trajectories are selected. Items participate in similarity matching; when the number of remaining historical trajectories is less than At that time, all remaining historical trajectories are used for similarity matching; weighted Euclidean distance is used. The calculation formula is ,in This is the current historical trajectory vector. For the first A historical degradation trajectory vector, The matrix is ​​a non-negative diagonal weight matrix. The diagonal elements are determined based on the reciprocal of the variance of each trajectory variable in the historical database, ensuring comparable contributions from variables of different scales in distance calculation. When the variance of a variable is zero, its corresponding weight is reset to zero, indicating that the variable does not participate in distance calculation in the current match. Similarity... The weighted Euclidean distance is mapped to a similarity value between 0 and 1 using an exponential kernel function. The calculation formula is as follows: ,in The preset similarity scale parameters, The determination of the initial similarity scale adopts a combination of preset initial values ​​and historical data updates; when the system is running for the first time or when the number of valid historical degradation trajectories in the historical bearing grease degradation trajectory database is insufficient, a preset initial similarity scale is used. ,in The condition that all current historical trajectory variables have been dimensionless is pre-set to 1; when the number of valid historical degradation trajectories reaches Then, the weighted Euclidean distance between each pair of historical trajectories is calculated to obtain the distance statistical distribution, and the data is updated based on the distance statistical distribution. Then, with the goal of minimizing the mean square error between the predicted bearing grease degradation index and the actual value, the updated index is... Perform one-dimensional search optimization; a large σ makes the similarity smoother with distance changes, while a small σ... The similarity is concentrated on the closest historical trajectory; the calculation of the bearing grease deterioration index uses a similarity-weighted average, and the calculation formula is as follows: , of which molecules express The sum of the bearing grease deterioration index corresponding to each historical trajectory at a future moment, weighted by current similarity; the higher the similarity, the greater the contribution of the future value of the corresponding historical trajectory to the prediction result; denominator In This represents the sum of all similarity weights. This is a preset, extremely small positive number used to prevent division by zero errors when the sum of similarities is zero. The method for determining this is to take the sum of the minimum non-zero similarities from the historical trajectory similarity calculation results and multiply it by a preset proportional coefficient; The preset future time step is set to match the time range within which the cold storage agent needs to intervene in advance, and is taken as an integer multiple of a single control cycle. When the sum of all similarities is lower than the preset effective similarity threshold, it indicates that the current historical trajectory differs significantly from any degradation pattern in the historical database. In this case, the prediction result is not reliable enough, and the weight of the predicted bearing grease degradation index in the subsequent predictive adjustment reward is reduced. When running for the first time and there is no valid historical degradation trajectory in the historical bearing grease degradation trajectory database, similarity matching prediction based on historical degradation trajectories is not performed. When the current bearing grease degradation index historical trajectory already has continuous valid data, the trend of the effective bearing grease degradation index change in the last two control cycles is used for short-term extrapolation, and the extrapolation result is limited to the range of 0 to 1 as the predicted bearing grease degradation index. When the number of historical bearing grease degradation trajectories reaches a certain threshold, the prediction is further adjusted accordingly. Then, switch to the historical degradation trajectory prediction method based on weighted Euclidean distance and similarity.

[0042] Each cold storage unit in the cold storage group is defined as a cold storage intelligent agent; the... Local observations of the cold storage intelligent agent include cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, evaporator fan speed, evaporator fan motor operating current, effective bearing grease degradation index, and predicted bearing grease degradation index; based on the natural logarithm of the ratio of the shared compressor unit's condensing pressure to the shared compressor unit's suction pressure, according to the formula... Calculate the shared refrigeration pressure state, where To share the condensing pressure of the compressor unit, To share the suction pressure of the compressor unit; based on the ratio of the total electrical power of the cold storage group to the reference value of the cold storage group's power, according to the formula... Calculate the dimensionless power state of the cold storage group, where This is a reference value for the power of cold storage groups. The total electrical power of the cold storage group is represented by the observations of each cold storage agent. The observations of the cold storage agents are combined with the local observations of each cold storage agent and the shared refrigeration pressure state, the dimensionless power state of the cold storage group, the average power of the demand sliding window, and the demand over-limit to form the observations of the cold storage agents. During the centralized training phase, the observations of each cold storage agent are combined into the global state.

[0043] It should be explained that each cold storage unit in the cold storage group is defined as an independent cold storage intelligent agent. The first cold storage agent corresponds to the first... A cold storage facility, wherein the intelligent agent makes decisions based solely on its own locally available observations and shared states that are allowed to be broadcast during operation; the first... Local observations by a cold storage agent are based on the cold storage temperature. Cold storage set temperature Cold storage compressor operating frequency Cold storage defrosting status Cold storage air door opening The fan speed of each evaporator in the cold storage Operating current of each evaporator fan motor Effective bearing grease deterioration index for each evaporator fan bearing and an index for predicting the degree of bearing grease deterioration The data are collected synchronously according to the control cycle from data acquisition, vibration characteristics and health benchmark calculation, and lubricating grease degradation assessment and prediction before entering local observation. For cases with multiple evaporator fans in the same cold storage, the speed, current, effective index, and predicted index of multiple fans can be aggregated into the local observation as averages. The refrigeration pressure status is shared. Condensing pressure from shared compressor unit With shared compressor unit suction pressure The ratio between them is obtained by taking the natural logarithm, and the calculation formula is: ,in and The absolute pressure is a positive value, expressed in kPa; the natural logarithmic transformation is used to map the pressure ratio to a symmetrical real number range, when the intake pressure decreases or the condensation pressure increases. An increase reflects an increase in the shared refrigeration pressure ratio. This decrease occurs when the suction pressure rises or the condensing pressure falls, providing a dimensionless shared pressure operating status quantity for each cold storage agent; the cold storage group also has a dimensionless power status. Based on the current total power of the cold storage group Divide by the reference value of cold storage group power The calculation formula is as follows: ,in The power reference value for the cold storage group is determined by taking one of the following three factors: the rated total installed power of the cold storage group, the historical maximum operating power, or the demand limit value. The demand limit value is preferred to ensure... It directly represents the utilization rate of current power relative to the upper limit of demand; The power consumption is measured and supplied in real time by the main incoming power metering device of the cold storage group; demand exceeding the limit. Using the aforementioned definition Calculation, where The average power of the demand sliding window. This is the demand limit value. A value greater than zero indicates that the average power of the current sliding window has exceeded the demand limit; the refrigeration pressure status is locally observed and shared by each cold storage agent. Dimensionless power state of cold storage group Average power of demand sliding window and demand exceeding the limit Together, we form a cold storage intelligent body observation system. During the formation process, continuous quantities in local observations are normalized and scaled according to their respective rated values. Binary defrosting conditions are preserved at their original values. The effective bearing grease degradation index and the predicted bearing grease degradation index directly utilize values ​​in the 0-1 range. During the intensive training phase, the observations from each cold storage agent are... The global state is constructed by assembling the data according to the fixed cold storage number sequence. The centralized commentator can use the global state to obtain joint operation information of the entire cold storage group and estimate the global value of the actions of each cold storage agent.

[0044] Each cold storage agent outputs the set temperature adjustment, refrigeration demand adjustment, and damper opening adjustment as continuous control actions, and outputs a defrost start request as a defrost control action. When multiple cold storage units share a common compressor unit, the refrigeration demand adjustment output by each cold storage agent is input into the shared refrigeration resource coordination mechanism. The shared refrigeration resource coordination mechanism determines the final operating command of the common compressor unit based on the refrigeration demand adjustment of each cold storage unit, the suction pressure of the common compressor unit, the condensing pressure of the common compressor unit, the common refrigeration capacity, the temperature safety margin of each cold storage unit, and the power status of the cold storage group. Action safety constraints are applied to the continuous control actions and defrost control actions. The action safety constraints include the upper and lower limits of the operating frequency of the cold storage compressor, the upper and lower limits of the cold storage damper opening, the upper and lower limits of the set temperature of the cold storage, the upper and lower limits of the predicted cold storage temperature, the pressure boundary of the common compressor unit, and the boundary of the common refrigeration capacity. When the predicted bearing grease deterioration index is greater than or equal to the warning threshold of the bearing grease deterioration index but less than the critical threshold of the bearing grease deterioration index, and the predicted cold storage temperature is close to the upper limit of the cold storage temperature, the action of reducing the refrigeration capacity is restricted.

[0045] It should be explained that the action outputs of each cold storage agent are generated by a policy network deployed in the cold storage controller. After receiving observations from the corresponding cold storage agent, the policy network outputs continuous control actions and defrosting control actions. The continuous control actions include adjusting the set temperature of the cold storage. Cold storage refrigeration demand adjustment volume Cold storage damper opening adjustment amount The cold storage refrigeration demand adjustment amount represents the adjustment amount of the cold storage compressor operating frequency in a cold storage with an independent compressor, and represents the percentage adjustment amount of the cold storage refrigeration demand in a cold storage with a shared compressor group; the defrost control action is based on a defrost start request. The output is in the form of a parameter. The request is converted into a binary action after threshold judgment. A defrost start request is generated when the continuous value output by the strategy network exceeds the preset defrost start judgment threshold; otherwise, the non-defrost state is maintained. The defrost start judgment threshold is determined based on the distribution of strategy outputs before and after normal defrost start in historical defrost operation data. When multiple cold storage units share a common compressor unit, the cold storage cooling demand adjustment amount output by each cold storage agent... To coordinate shared refrigeration resources, the first step is to convert the refrigeration demand adjustment of each cold storage facility into an increase in refrigeration capacity demand, and then combine this with the suction pressure of the shared compressor unit. Condensing pressure The maximum refrigeration capacity that can be allocated to each cold storage unit is determined based on the current shared refrigeration capacity, which is obtained by interpolation of the performance curve of the shared compressor unit under the current suction pressure, condensing pressure, and ambient temperature conditions. The shared refrigeration resource coordination calculates the temperature safety margin of each cold storage unit during allocation. ,in For the first The upper limit of the temperature of each cold storage room: When the total cooling demand of all cold storage rooms does not exceed the shared cooling capacity, the cooling capacity is allocated according to the proportion of the cooling demand adjustment of each cold storage room. When the total exceeds the shared cooling capacity, priority is given to allocating cooling capacity to cold storage rooms with smaller temperature safety margins. (The last sentence appears to be incomplete and possibly refers to a different context.) If the power exceeds a preset power coordination threshold, the refrigeration capacity of non-critical cold storage facilities will be reduced during off-peak hours, ultimately generating the final operating command for the shared compressor unit; the preset power coordination threshold is based on the dimensionless power status of the cold storage group during its historical normal operation. It has been determined that historical normal operation data shows no instances of demand exceeding limits and cold storage temperatures meeting operational requirements. The samples are arranged in ascending order. A boundary value is determined based on the power range required to maintain temperature safety and ensure normal operation of shared refrigeration equipment during the high-power operation phase of the cold storage group. This boundary value is used as a preset power coordination threshold to trigger refrigeration load staggering in advance when the cold storage group enters the high-power operation range. Action safety constraints are used to correct the continuous control actions and defrosting control actions output by the strategy network item by item. The upper and lower limits of the cold storage compressor operating frequency are determined based on the compressor nameplate parameters and the inverter's allowable range. The upper and lower limits of the cold storage damper opening are determined based on the damper actuator's stroke range. The upper and lower limits of the cold storage set temperature are determined based on the cold storage process requirements. The upper and lower limits of the predicted cold storage temperature are determined based on the cold storage temperature safety requirements. The cold storage thermal state prediction model is calculated based on the current cold storage temperature, the operating status of the refrigeration equipment, candidate control actions, and the cold storage heat load. The model can be updated online after parameter identification using a first-order inertial model. The shared compressor unit pressure boundary and shared refrigeration capacity boundary are determined based on the equipment design pressure limit and the maximum refrigeration capacity under current environmental conditions. When predicting the bearing grease deterioration index... Greater than or equal to the bearing grease deterioration index warning threshold And less than the critical threshold of the bearing grease deterioration index. Meanwhile, it is predicted that the cold storage temperature will meet the requirements. When the predicted cold storage temperature is close to the upper limit of the cold storage temperature range, it is determined that... The preset safety margin threshold for cold storage temperature is defined as follows: the threshold is based on the allowable deviation of cold storage temperature. The bearing grease deterioration index warning threshold is determined by multiplying by a preset proportional coefficient. Based on the bearing grease degradation index distribution of confirmed normal and warning state bearing grease samples, the boundary value between normal and warning states is determined under the condition of minimizing the misclassification rate of samples in normal and warning states; the critical threshold of the bearing grease degradation index is defined. Based on the bearing grease degradation index distribution of confirmed early warning and critical state bearing grease samples, the method is determined to minimize the misclassification rate of early warning and critical state samples while satisfying the following conditions. The boundary value between the warning state and the critical state is determined under the condition that the preset proportional coefficient is based on the prediction error during the cold storage temperature control process relative to the allowable deviation of the cold storage temperature. Based on the historical distribution, the proportion that can cover the main fluctuation range of the prediction error while maintaining a sufficient temperature safety margin is determined as the preset proportional coefficient. This ensures that the cold storage temperature safety margin threshold can cover the prediction error and trigger a reduction in refrigeration capacity before the predicted cold storage temperature approaches its upper limit. Under these conditions, the action safety constraint restricts the reduction in refrigeration capacity, specifically by cutting off the negative portion of the cold storage refrigeration demand adjustment to zero, cutting off the negative portion of the cold storage damper opening adjustment to zero, and restricting the positive increase of the cold storage set temperature adjustment.

[0046] The control actions, after being constrained by safety rules, are sent to the cold storage actuators. Upon execution, new cold storage temperatures, set temperatures, operating status of the refrigeration equipment, defrosting status, damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, total power of the cold storage group, and new bearing vibration acceleration data are generated. The reinforcement learning reward is calculated based on the new total power of the cold storage group, the average power of the demand sliding window, the cold storage temperature, the predicted bearing grease degradation index, and the control action amplitude. The reinforcement learning reward is calculated according to the formula... Calculations, including the item requiring excess quantity. , , This represents the dimensionless power state of the cold storage group. , , Non-negative reward weights To predictively adjust the reward, new operating data and vibration acceleration data are used as inputs for the next control cycle.

[0047] It needs to be explained that, after the action safety constraints are applied, the continuous control actions and defrosting control actions are sent to the cold storage actuators respectively. Specifically, the cold storage set temperature adjustment is sent to the cold storage temperature controller; the cold storage refrigeration demand adjustment is converted into a compressor frequency command by the compressor controller and sent to the compressor inverter; the cold storage damper opening adjustment is sent to the damper actuator; and the defrosting start request is sent to the defrosting control. After the control actions are executed, the internal thermal environment of the cold storage, the operating status of the refrigeration equipment, and the shared refrigeration status change accordingly, resulting in new cold storage temperature, cold storage set temperature, cold storage refrigeration equipment operating status, cold storage defrosting status, and cold storage damper opening. Simultaneously, the shared compressor unit's suction pressure and condensing pressure stabilize to new values ​​after the refrigeration load change. The total incoming power metering device for the cold storage group updates the total power of the cold storage group, and the vibration acceleration sensor continues to collect and generate new bearing vibration acceleration data. In the reinforcement learning reward calculation, the dimensionless power state of the cold storage group... Using the current total electrical power of the cold storage group Divide by the reference value of cold storage group power get, Take the demand limit value of the cold storage group; items exceeding the demand limit. Defined as ,in The average power of the demand sliding window. The demand limit value is preset by the power supply contract; the normalized variable for cold storage temperature deviation. Defined as ,in For the first The temperature of the cold storage room For the first Each cold storage unit is set to a specific temperature. For the first The allowable temperature deviation for each cold storage room is taken as the corresponding value for the required temperature control accuracy; reward formula. middle, This indicates a penalty for the power utilization level of the cold storage group, when The penalty increases when the value deviates from zero, and extreme penalty values ​​are limited through exponential saturation. This represents the penalty for exceeding the demand limit; when the average power of the sliding window exceeds the demand limit, i.e. When greater than zero Increase; This represents the penalty applied to the average temperature deviation across all cold storage facilities; the larger the temperature deviation, the greater the penalty. The closer ; , , The non-negative reward weights are obtained by constructing a pairwise comparison matrix based on the priority of energy management objectives for the cold storage group using the analytic hierarchy process (AHP). The relative importance of each pair of objectives (energy saving, demand management, and temperature safety) is used as matrix elements. The normalized eigenvector corresponding to the largest eigenvalue of the matrix is ​​calculated to obtain the initial weights. During multi-agent reinforcement learning training, the initial weights are fine-tuned based on the optimal cumulative reward of the policy. Predictive adjustment of rewards is also performed. The calculation is based on the increase in the predicted bearing grease deterioration index relative to the current effective bearing grease deterioration index, the safety margin between the predicted cold storage temperature and the upper limit of the cold storage temperature, and the control action range. Specifically: ,in, The preset predictive adjustment reward weights are determined by maximizing the cumulative reward during offline reinforcement learning training, based on the relative importance between the predictive equipment health adjustment target and the energy-saving control target of the cold storage group. , To predict the degree of deterioration of bearing grease, This is an index representing the current effective degree of deterioration of bearing grease. For the first The upper limit of the storage temperature of a cold storage unit. To predict the cold storage temperature after executing candidate control actions, To preset the safety margin threshold for cold storage temperature; the continuous control actions are dimensionless according to their respective allowable adjustment ranges before calculating the predictive adjustment reward; new cold storage temperature, cold storage set temperature, cold storage refrigeration equipment operating status, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, total power of the cold storage group, and new bearing vibration acceleration data are all fed back to data acquisition and vibration characteristic and health benchmark calculation at the beginning of the next control cycle, and the complete process of vibration signal processing, healthy vibration reference benchmark construction, operating condition normalization, mechanical defect suppression, bearing grease deterioration index calculation, prediction index calculation, cold storage agent state construction, and multi-agent decision-making is re-executed, forming a closed loop of collaborative energy-saving control of the cold storage group; in the multi-agent reinforcement learning training stage, the MAPPO algorithm of multi-agent near-end policy optimization is adopted with centralized training and decentralized execution; each cold storage agent is configured with a policy network to correspond to the cold storage agent observations. To input and output the continuous control actions and defrosting control actions, a centralized value network is used to receive the global state formed by the combination of observations from each cold storage agent. According to the reinforcement learning reward The evaluation focuses on the joint control results of each cold storage agent. During training, each agent generates actions based on its current strategy, which are then executed after coordination with shared refrigeration resources and action safety constraints. This process yields corresponding rewards and the global state for the next control cycle, forming training samples used to update the policy networks of each agent and the centralized value network. Through iterative interactions of state, action, reward, and next state, each agent gradually learns a collaborative control strategy under the combined influence of equipment health status, cold storage temperature, shared refrigeration capacity, cold storage group power status, and demand constraints. After training, the policy networks of each agent are deployed in their respective cold storage controllers. During online operation, these controllers perform forward inference based on current agent observations and output control actions, without updating policy parameters online.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning, characterized in that, include: The cold storage group data acquisition and demand calculation module is used to collect data such as bearing vibration acceleration, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, and total power of the cold storage group, and calculate the average power of the demand sliding window. The vibration characteristics and health benchmark calculation module is used to perform high-frequency bandpass filtering, envelope demodulation, and envelope spectrum calculation on bearing vibration acceleration data. It dynamically calculates the characteristic frequency of the target bearing based on the evaporator fan speed and extracts the characteristic frequency amplitude of the target bearing. The representative characteristic amplitude and representative characteristic energy index are obtained by root mean square averaging. Calculate the healthy vibration reference benchmark based on the healthy operation sample and the current operating conditions, and calculate the normalized characteristic energy ratio and mechanical defect suppression weight of the operating conditions. The grease degradation assessment and prediction module is used to calculate the degradation contribution and comprehensive degradation amount based on the normalized characteristic energy ratio of the working condition and the mechanical defect suppression weight, construct the bearing grease degradation degree index and perform validity processing; and obtain the predicted bearing grease degradation degree index by similarity matching between the historical trajectory of the effective bearing grease degradation degree index and the historical bearing grease degradation trajectory. The multi-agent collaborative decision-making and safety constraint module is used to form the cold storage agent status by combining the effective bearing grease deterioration index, the predicted bearing grease deterioration index, the cold storage operation status, the evaporator fan operation status, the shared refrigeration status, and the cold storage group power status. Each cold storage agent outputs continuous control actions and defrosting control actions based on the cold storage agent status to coordinate shared refrigeration resources and enforce action safety constraints. The control execution and status update module is used to execute control actions after the action safety constraints have been met, update the cold storage operation status and the energy status of the cold storage group, and form the input for the next control cycle.

2. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The data collected includes the bearing vibration acceleration of each evaporator fan in the cold storage group, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, and total power of the cold storage group. The calculation of the demand sliding window average power involves: feedback from vibration acceleration sensors, speed sensors or frequency converters installed on the evaporator fan bearing housings, current sensors, cold storage temperature sensors, cold storage controllers, damper position feedback, and shared compressor unit suction pressure. Side pressure sensors, shared compressor unit exhaust side pressure sensors, and cold storage group main power metering devices collect data on bearing vibration acceleration, evaporator fan speed, evaporator fan motor operating current, cold storage temperature, cold storage set temperature and defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, and total power of the cold storage group, respectively. The average power of the demand sliding window is calculated based on the total power of the cold storage group, which is the average value of the total power of the cold storage group within the preset demand window length. All collected data are synchronized according to a unified timestamp.

3. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The process of performing high-frequency bandpass filtering, envelope demodulation, and envelope spectrum calculation on bearing vibration acceleration data, dynamically calculating the target bearing characteristic frequency based on the evaporator fan speed, extracting the target bearing characteristic frequency amplitude, and obtaining representative characteristic amplitude and representative characteristic energy index through root mean square averaging includes: continuously sampling the bearing vibration acceleration data and dividing it into segments according to a preset vibration sampling frame length. A series of vibration sampling frames, The frequency band of the high-frequency impact signal is determined based on the evaporator fan bearing structure, the frequency response range of the vibration acceleration sensor, and the sampling frequency. Bandpass filtering is performed on each vibration sampling frame to obtain the high-frequency impact signal. An analytical signal is calculated from the high-frequency impact signal, and the modulus of the analytical signal is used as the envelope signal. A discrete Fourier transform is performed on the envelope signal to obtain the envelope spectrum. The characteristic frequency of the target bearing is dynamically calculated based on the evaporator fan speed and bearing structural parameters, including the number of rolling elements, rolling element diameter, pitch circle diameter, and contact angle. The amplitude of the characteristic frequency of the target bearing is extracted from the envelope spectrum. Within a preset sliding time window, the characteristic frequency of the same target bearing is... vibration sampling frame amplitude Perform the root mean square average according to the formula. Calculate the representative characteristic amplitude; obtain the representative characteristic energy index based on the square of the representative characteristic amplitude. .

4. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The calculation of the healthy vibration reference benchmark based on healthy operation samples and current operating conditions includes: pre-establishing a healthy operation sample database to store operating data when the bearing grease is in a normal state and there are no obvious mechanical defects; the mechanical defects include damage to the bearing outer ring, damage to the bearing inner ring, and damage to the rolling elements; selecting evaporator fan speed, evaporator fan motor operating current, the difference between the cold storage temperature and the cold storage set temperature, cold storage damper opening, shared compressor unit suction pressure, and shared compressor unit condensing pressure as corresponding reference values ​​from the healthy operation sample database; dividing the six current operating parameters by their corresponding reference values ​​to obtain six dimensionless variables for operating conditions, which are uniformly denoted as... Representative feature amplitudes and corresponding dimensionless variables of operating conditions are extracted from the healthy operation sample database to form healthy operation samples. Based on the healthy operation samples, the relationship between the dimensionless variables of centralized operating conditions and the representative feature amplitudes is fitted, and centralized operating condition variables are defined. To obtain a healthy vibration reference standard The formula is: ,in, When the dimensionless variables of the six operating conditions are all taken as their respective health reference values, the first... The healthy reference amplitude corresponding to the characteristic frequency of each target bearing For the first The characteristic frequency of the target bearing corresponds to the first... Parameters of a linear term model with dimensionless variables under centralized operating conditions. For the first The characteristic frequency of the target bearing corresponds to the first... The and the first Model parameters for interaction terms between dimensionless variables in a centralized operating condition. The first of the six dimensionless variables for operating conditions One variable, For the corresponding centralized operating conditions, dimensionless variables are used. The first of the six dimensionless variables for operating conditions One variable, For the corresponding centralized operating conditions, there are dimensionless variables.

5. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The calculation of the normalized characteristic energy ratio and mechanical defect suppression weight includes: [calculating the representative characteristic amplitude]. Compared with healthy vibration reference standards Normalization is performed to obtain the normalized characteristic energy ratio under the operating condition. The formula is: ,in The value is a preset, extremely small positive number with dimensions consistent with the representative characteristic amplitude; the concentration of mechanical defect features is calculated based on the representative characteristic energy index. The concentration degree of mechanical defect features is equal to the ratio of the sum of the representative characteristic energy indices corresponding to the mechanical defect feature frequencies to the sum of the representative characteristic energy indices corresponding to the feature frequencies of all target bearings. The set of mechanical defect feature frequencies is composed of the target bearing feature frequencies corresponding to the mechanical defects. Mechanical defect suppression weights are calculated based on the concentration degree of mechanical defect features. The formula is: ,in This is the preset mechanical defect suppression coefficient; when the concentration of mechanical defect features is continuous When a control cycle is greater than or equal to a preset mechanical defect judgment threshold, the evaporator fan is determined to be in an abnormal mechanical defect state. It is a preset positive integer.

6. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The process of calculating the degradation contribution and overall degradation amount based on the normalized characteristic energy ratio under operating conditions and the mechanical defect suppression weight, constructing a bearing grease degradation index, and performing effectiveness processing includes: calculating the degradation contribution of each target bearing characteristic frequency based on the normalized characteristic energy ratio under operating conditions and the mechanical defect suppression weight. The formula is: ,in For the first Nonlinear sensitivity index of the characteristic frequency of a target bearing ; Calculate the overall degradation amount The formula is: ;in , The first The and the first The characteristic frequency weights of each target bearing are all non-negative and sum to 1. These are the nonlinear interaction coefficients between features. For the first The degradation contribution of each target bearing characteristic frequency; candidate values ​​for the bearing grease degradation index are calculated based on the overall degradation amount. The formula is: ,in The preset global nonlinear index is used; when the evaporator fan is not in a mechanical defect abnormal state, the candidate value of the bearing grease deterioration index is determined as the effective bearing grease deterioration index; when the evaporator fan is in a mechanical defect abnormal state, the effective bearing grease deterioration index of the previous moment and the candidate value of the current bearing grease deterioration index are calculated according to the formula. A smooth update was performed to obtain an effective bearing grease degradation index; among which... For smoothing coefficients, This represents the index of the degree of deterioration of the effective bearing grease at the previous moment.

7. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The method of obtaining a predicted bearing grease degradation index by similarity matching between the historical trajectory of the effective bearing grease degradation index and the historical bearing grease degradation trajectory includes: continuously recording the degradation index of each evaporator fan bearing. The effective bearing grease degradation index at each time point was recorded simultaneously with the above. The dimensionless variables of the operating conditions corresponding to each time point are used to form the historical trajectory of the current bearing grease degradation index. A historical bearing grease degradation trajectory database is pre-established to store the historical trajectory of the bearing grease degradation index during the gradual degradation process, the corresponding historical trajectory of the operating conditions, and the bearing grease degradation index after a preset future time. H historical bearing grease degradation trajectories are obtained from the historical bearing grease degradation trajectory database, where H is a preset positive integer. Each historical bearing grease degradation trajectory contains the same variables, time length, and normalization method as the current bearing grease degradation index historical trajectory. The weighted Euclidean distance between the current bearing grease degradation index historical trajectory and each historical bearing grease degradation trajectory is calculated. Similarity is calculated based on weighted Euclidean distance. ,in The similarity scale parameter is set as a preset value; the bearing grease deterioration index corresponding to each historical trajectory after a preset future time is weighted and averaged according to similarity to obtain the predicted bearing grease deterioration index. The formula is: ,in For the preset future time step, For the current moment The subsequent predicted time, It is a preset minimum positive number.

8. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The process of forming a cold storage intelligent agent status by considering the effective bearing grease degradation index, the predicted bearing grease degradation index, the cold storage operating status, the evaporator fan operating status, the shared refrigeration status, and the cold storage group power status includes: defining each cold storage unit in the cold storage group as a cold storage intelligent agent; the first... Local observations of the cold storage intelligent agent include cold storage temperature, cold storage set temperature, cold storage compressor operating frequency, cold storage defrosting status, cold storage damper opening, evaporator fan speed, evaporator fan motor operating current, effective bearing grease degradation index, and predicted bearing grease degradation index; based on the natural logarithm of the ratio of the shared compressor unit's condensing pressure to the shared compressor unit's suction pressure, according to the formula... Calculate the shared cooling pressure state, where To share the condensing pressure of the compressor unit, To share the suction pressure of the compressor unit; based on the ratio of the total electrical power of the cold storage group to the reference value of the cold storage group's power, according to the formula... Calculate the dimensionless power state of the cold storage group, where This is a reference value for the power of cold storage groups. The total electrical power of the cold storage group is represented by the observations of each cold storage agent. The observations of the cold storage agents are combined with the local observations of each cold storage agent and the shared refrigeration pressure state, the dimensionless power state of the cold storage group, the average power of the demand sliding window, and the demand over-limit to form the observations of the cold storage agents. During the centralized training phase, the observations of each cold storage agent are combined into the global state.

9. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, Each cold storage intelligent agent outputs continuous control actions and defrost control actions based on its status, coordinating shared refrigeration resources and constraining action safety. This includes: each cold storage intelligent agent outputting the cold storage set temperature adjustment, cold storage refrigeration demand adjustment, and cold storage damper opening adjustment as continuous control actions, and outputting a defrost start request as a defrost control action; when multiple cold storages share a common compressor unit, the cold storage refrigeration demand adjustment output by each cold storage intelligent agent is input into the shared refrigeration resource coordination, which is based on the refrigeration demand adjustment of each cold storage, the suction pressure of the common compressor unit, the condensing pressure of the common compressor unit, and the common refrigeration capacity. The final operating command of the shared compressor unit is determined by the quantity, temperature safety margin of each cold storage room, and power status of the cold storage group; action safety constraints are implemented for continuous control actions and defrosting control actions, including the upper and lower limits of cold storage compressor operating frequency, cold storage damper opening, cold storage set temperature, predicted cold storage temperature, shared compressor unit pressure boundary, and shared refrigeration capacity boundary; when the predicted bearing grease deterioration index is greater than or equal to the bearing grease deterioration index warning threshold but less than the bearing grease deterioration index critical threshold, and the predicted cold storage temperature is close to the upper limit of the cold storage temperature, the action of reducing refrigeration capacity is restricted.

10. The collaborative energy-saving control system for cold storage clusters based on multi-agent reinforcement learning according to claim 1, characterized in that, The execution of control actions after safety constraints is performed updates the cold storage operating status and the energy status of the cold storage group, forming the input for the next control cycle. This includes: sending the control actions after safety constraints to the cold storage actuators; generating new cold storage temperature, cold storage set temperature, cold storage refrigeration equipment operating status, cold storage defrosting status, cold storage damper opening, shared compressor unit suction pressure, shared compressor unit condensing pressure, total power of the cold storage group, and new bearing vibration acceleration data after the control actions are executed; calculating reinforcement learning rewards based on the new total power of the cold storage group, average power of the demand sliding window, cold storage temperature, predicted bearing grease degradation index, and control action amplitude, with the reinforcement learning rewards calculated according to the formula. Calculations, including the requirement exceeding the limit item. , , This represents the dimensionless power state of the cold storage group. , , Non-negative reward weights To predictively adjust the reward; new operating data and vibration acceleration data are used as inputs for the next control cycle.

Images