A big data-based heating system fault positioning system and method

Abstract

The present application relates to the technical field of fault detection, in particular to a heating system fault positioning system and method based on big data, the system comprising: a multi-source sensing module, a disturbance feature module, a path modeling module, a frequency spectrum matching module, and a fault positioning module.In the present application, real-time disturbance sequences are constructed through time window segmentation and parameter offset calculation, the separation of active response chains and abnormal propagation paths is achieved through directional joint state vectors and topological relationship tables, the dynamic time warping algorithm is used to eliminate the time asynchrony of multi-device signal transmission, the real-time parameter bidirectional verification mechanism of frequency domain main frequency band energy markers, valve opening degrees, and pump rotation speeds is combined to solve the path confusion problem of traditional methods in multi-node parameter coupling scenarios, improve the tracing efficiency of concurrent faults in complex pipe network systems, overcome the adaptability limitations of single-dimensional threshold mechanisms to device performance degradation, and reduce the probability of false positioning caused by signal delay superposition.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a HVAC system fault location system and method based on big data. Background Art

[0002] The field of fault detection technology primarily involves the technical means of identifying, diagnosing, and locating abnormal conditions that occur during the operation of various engineering systems, mechanical equipment, or electronic devices. The core of this technical field includes fault signal acquisition, fault identification methods, fault diagnosis model construction, fault source tracing, and fault characteristic parameter extraction. It is widely used in various fields, including industrial manufacturing, power systems, transportation, aerospace, and building automation. Fault detection technology is particularly critical in HVAC systems, which have complex system structures and volatile operating environments. Common faults such as abnormal chilled water system circulation, imbalanced temperature and humidity control in air handling units, and reduced energy efficiency in refrigeration equipment all rely on accurate fault detection mechanisms for timely response and location.

[0003] Traditional HVAC system fault location systems are used to identify and locate abnormal operating conditions during the operation of heating, ventilation, and air conditioning systems. These systems primarily address technical issues such as unstable equipment operation, abnormal energy consumption, and temperature and humidity fluctuations within the HVAC system. Traditional HVAC system fault location utilizes temperature sensor data recording, flow monitoring data collection, and a rule-based judgment based on operation logs. By performing attribution analysis on specific data features such as temperature trends, uneven water flow distribution, or energy consumption fluctuations, these systems can identify abnormal behavior of specific components, such as clogged fan coil units, abnormal pump operation, or stuck valves.

[0004] Traditional HVAC fault location uses a single sensor threshold and linear data attribution model. When multiple pumps in the chilled water system operate in parallel, it is unable to analyze the cross-device correlation characteristics of temperature and pressure parameters. When abnormal signals cross-propagate in the pipeline network, the judgment rules based on independent parameter fluctuations cannot distinguish between normal operating conditions and real fault triggers. For example, when the temperature control anomaly of the air handling unit and the fan speed fluctuation are superimposed, the static threshold mechanism can easily misjudge the related parameter fluctuations as independent fault points. At the same time, it lacks the ability to dynamically compensate for parameter drift during equipment performance degradation, resulting in hidden faults in long-term operating systems being difficult to capture in a timely manner. Summary of the Invention

[0005] In order to solve the problem that traditional HVAC fault location uses a single sensor threshold and linear data attribution model, when multiple pumps in the chilled water system are running in parallel, it is impossible to analyze the cross-device correlation characteristics of temperature and pressure parameters. When abnormal signals have cross-path propagation in the pipe network, the judgment rules based on independent parameter fluctuations are difficult to distinguish between normal operating condition adjustment and real fault triggering. For example, when the temperature control anomaly of the air handling unit is superimposed with the fan speed fluctuation, the static threshold mechanism is prone to misjudge the associated parameter fluctuation as an independent fault point. At the same time, it lacks the ability to dynamically compensate for parameter drift during the attenuation of equipment performance, resulting in the technical problem that hidden faults of long-term running systems are difficult to be captured in time, the embodiment of the present invention provides a HVAC system fault location system and method based on big data. The technical solution is as follows:

[0006] On the one hand, a HVAC system fault location system based on big data is provided, which includes:

[0007] The multi-source sensing module is used to collect flow rate, pressure, and temperature through a sensor group, split the data into time windows, calculate parameter offsets, generate real-time disturbance sequences, and pass them to the disturbance feature module;

[0008] A disturbance feature module is used to calculate the absolute value of the real-time disturbance sequence, compare it with the fluctuation threshold, mark the effective disturbance point, extract the direction and combine it into a joint state vector, and pass it to the path modeling module;

[0009] A path modeling module is used to call the joint state vector, determine the directional consistency of adjacent nodes based on the topological relationship table, divide the active response chain into an abnormal response chain, align the response delay using a dynamic time warping algorithm, output a directional propagation path set, and pass it to the spectrum matching module;

[0010] A spectrum matching module is used to extract raw data by calling the timestamp of the directional propagation path set, perform Fourier transform to extract the main frequency band, compare the energy to mark the abnormal frequency band, output the frequency response correlation path, and transmit it to the fault location module;

[0011] The fault location module is used to retrieve the device mapping table to obtain the component number according to the starting node of the frequency response association path, verify the real-time parameter deviation of the valve opening and the pump speed, and output the fault component number if the deviation exceeds the tolerance threshold.

[0012] As a further solution of the present invention, the real-time disturbance sequence specifically includes flow rate offset, pressure offset, and temperature offset; the joint state vector includes a valid disturbance coordinate set, a disturbance direction characteristic matrix, and a normalized fluctuation amplitude; the directional propagation path set specifically refers to an active response topology chain, an abnormal response topology chain, and a delay compensation coefficient; the frequency response association path includes a main frequency band amplitude spectrum, an abnormal frequency band energy distribution, and a phase alignment parameter; and the fault component number specifically includes a valve opening deviation value, a pump speed deviation value, and a safe operation threshold.

[0013] As a further solution of the present invention, the multi-source sensing module includes:

[0014] The multi-source sensor acquisition submodule collects flow pulse signals, pressure voltage signals, and temperature resistance signals, performs analog-to-digital conversion and fixed-time interception, performs multi-sensor timing alignment, and generates a multi-source time window data set;

[0015] The dynamic offset calculation submodule calls the multi-source time window data set, calculates the absolute value of the flow velocity offset at the beginning and end, analyzes the pressure window standard deviation, extracts the temperature linear trend slope, and sums the three parameter offsets weighted by the range ratio to generate a dynamic offset coefficient set;

[0016] The disturbance sequence generation submodule fits the displacement coefficient change curve based on the dynamic displacement coefficient set, calculates the rate of change using a sliding window, compares the original percentile threshold, and generates a real-time disturbance sequence when the threshold is exceeded three times in a row.

[0017] As a further solution of the present invention, the disturbance feature module includes:

[0018] The disturbance intensity quantification submodule reads the disturbance absolute sequence, performs intensity marking processing, compares the preset dynamic fluctuation threshold point by point, calculates the percentage of points exceeding the threshold, and generates a disturbance intensity coefficient;

[0019] The fluctuation threshold setting method is based on the range of the mean plus or minus two standard deviations of the original parameter data. It is obtained by analyzing real-time operation data and experimentally verifying 10 HVAC equipment. The 10 HVAC equipment covers three typical models. The sampling frequency is 100Hz and the confidence interval is set to 95%.

[0020] The effective disturbance identification submodule calculates the dynamic threshold adjustment factor based on the disturbance intensity coefficient, marks the coordinates of the points exceeding the threshold, extracts the positive and negative polarity features using a sign function, constructs a polarity-coordinate mapping table with a timestamp, and generates a valid disturbance coordinate set;

[0021] The state vector synthesis submodule calls the polarity characteristics of the effective perturbation coordinate set, normalizes the amplitude of the positive and negative polarities by the range method, and performs tensor splicing of the normalized amplitude and the original coordinates in time sequence to generate a joint state vector.

[0022] As a further solution of the present invention, the dynamic threshold adjustment factor is calculated using the formula:

[0023] ;

[0024] in, Represents the percentage adjustment factor of the dynamic fluctuation threshold at time t, which is a dimensionless parameter. Represents the weight coefficient of the disturbance intensity coefficient, which is a dimensionless parameter. Represents the perturbation intensity coefficient generated by the previous step, which is a dimensionless parameter. Represents the adjustment factor of the gradient change at the current moment, in s / Pa. Represents the discrete gradient value of the disturbance sequence at time t, in Pa / s, Represents the original data window length, Represents the attenuation weight of the i-th original window, in units of 1 / s, Represents the nonlinear activation coefficient, which is a dimensionless parameter. Represents the integrated value of the disturbance energy in the i-th window, in Pa 2 ·s.

[0025] As a further solution of the present invention, the path modeling module includes:

[0026] The directional consistency judgment submodule calls the connection relationship between the node displacement parameters and the topological relationship table in the joint state vector, extracts the displacement components of the node on the X-axis and Y-axis to construct a two-dimensional displacement vector, calculates the cosine value of the displacement direction angle between adjacent nodes according to the vector dot product formula, compares the cosine value with the directional consistency threshold item by item, and generates the node offset coefficient;

[0027] The node offset coefficient represents the intensity of the node response trend change and serves as a reference for chain division;

[0028] The response chain division submodule establishes a node status classification matrix based on the node offset coefficient, divides the coefficient distribution into quartile intervals, identifies nodes in the upper quartile and lower quartile and classifies them into active chains and abnormal chains, counts the spatial coordinates of the nodes in the chain, and generates a chain identification matrix;

[0029] The delay alignment submodule calls the chain identification matrix to extract the time series of the active response chain and the abnormal response chain, constructs the cost matrix of the dynamic time warping algorithm, recursively searches the cumulative path alignment waveform, uses cubic spline interpolation to fill in the gaps, and generates a set of directed propagation paths.

[0030] As a further solution of the present invention, the spectrum matching module includes:

[0031] The time series data acquisition submodule obtains the time stamp of the directional propagation path set, calls the original signal data of the corresponding time interval, detects the sampling rate matching, performs interpolation compensation on the offset time series segment, and generates the original signal time series set;

[0032] The frequency domain conversion submodule performs Fourier transform based on the original signal time series set, calculates the frequency band spectrum amplitude, extracts the main frequency band parameters, synchronously calculates the original normal working condition energy distribution characteristics, and establishes the main frequency band energy feature set;

[0033] The frequency band anomaly detection submodule calls the main frequency band energy feature set, separates the current energy spectrum from the reference energy distribution, calculates the frequency band energy difference, marks the abnormal frequency band three-dimensional parameters, integrates the abnormal frequency band mark set, and generates a frequency response correlation path.

[0034] As a further solution of the present invention, the frequency spectrum amplitude of the frequency band is calculated using the formula:

[0035] ;

[0036] in, represents the spectrum amplitude of the kth frequency band, Represents the energy value of the k+mth frequency point, which comes from the Fourier transform result of the original signal time series set, and the unit is J. Representative The average value of the reference energy of three adjacent frequency points in the frequency band, in J. Represents the adjacent frequency weight factor , Represents the energy value of the k+nth frequency point after Fourier transform of the original signal time series set, the unit is J, Representative Normalized vibration phase difference of the frequency band, Representative Frequency band power change rate, unit is J / s, System characteristic response time, in seconds.

[0037] As a further solution of the present invention, the fault location module includes:

[0038] The component association retrieval submodule obtains the starting node number of the frequency response association path, matches the relationship between the node and the component number, reads the unique identifier corresponding to the current node, and generates a component association number;

[0039] The real-time parameter acquisition submodule reads the valve sensor pulse voltage signal and the pump body frequency current signal based on the component association number, converts them into standard opening percentage and standard speed range, and generates dual-channel dynamic parameters;

[0040] The deviation threshold determination submodule calls the dual-channel dynamic parameters, calculates the opening difference and the speed change rate, synchronously compares the system tolerance threshold, and outputs the fault component number;

[0041] The tolerance threshold is based on a 5% fluctuation in the upper limit of the parameter rating. After verification by 40 dynamic loadings in the laboratory and 3 days of continuous operation in the production environment, the error rate is less than ±0.8%.

[0042] On the other hand, a method for locating HVAC system faults based on big data is provided. The method is performed based on the above-mentioned HVAC system fault locating system based on big data and includes the following steps:

[0043] S1: Flow rate, pressure, and temperature values ​​are collected through a multi-channel sensor group. The sliding window algorithm is used to segment the continuous time series data into equal lengths. The offset between the parameters in the window and the reference value is calculated to generate a real-time disturbance sequence.

[0044] S2: calling the real-time disturbance sequence, inputting the data points into the absolute value operation function for amplitude conversion, screening the effective disturbance points through the preset fluctuation threshold, extracting the change direction of adjacent disturbance points, and combining the joint state vector;

[0045] S3: Based on the node topology relationship of the joint state vector, a directional consistency determination algorithm is used to divide the response chain types, and the active response chain and the abnormal response chain are input into the dynamic time warping algorithm for delay alignment, and a directional propagation path set is output;

[0046] S4: calling the timestamp index of the directed propagation path set, extracting the corresponding original vibration data, calculating the spectrum distribution through fast Fourier transform, comparing the energy ratio threshold of the standard frequency band, marking the abnormal frequency band, and generating the frequency response correlation path;

[0047] S5: According to the starting node number of the frequency response association path, the device topology database is retrieved to obtain the corresponding component code, and the valve opening setting value and the pump speed feedback value are compared in real time. When the deviation exceeds the preset tolerance, the fault component number is output.

[0048] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0049] A real-time disturbance sequence is constructed based on time window segmentation and parameter offset calculation. The active response chain and the abnormal propagation path are separated through the directional joint state vector and the topological relationship table. The dynamic time warping algorithm is used to eliminate the temporal asynchrony of multi-device signal transmission. The two-way verification mechanism of the real-time parameters of the main frequency band energy mark in the frequency domain and the valve opening and pump speed is combined to solve the path confusion problem of the traditional method in the multi-node parameter coupling scenario, improve the tracing efficiency of concurrent faults in complex pipe network systems, overcome the adaptability limitations of the single-dimensional threshold mechanism to equipment performance degradation, and reduce the probability of mislocation caused by signal delay superposition. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0051] Figure 1 It is a system flow chart of the present invention;

[0052] Figure 2 Schematic diagram of the system framework of the present invention;

[0053] Figure 3 It is a schematic diagram of the method steps of the present invention. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0055] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0056] In the embodiments of the present invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same. The terms "of," "corresponding," and "corresponding" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same.

[0057] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0058] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0059] The embodiment of the present invention provides a HVAC system fault location system based on big data, such as Figure 1 The flow chart of the HVAC system fault location system based on big data is shown, and the system includes:

[0060] The multi-source sensing module is used to collect flow rate, pressure, and temperature through a sensor group, split the data into time windows, calculate parameter offsets, generate real-time disturbance sequences, and pass them to the disturbance feature module;

[0061] The disturbance feature module is used to calculate the absolute value of the real-time disturbance sequence, compare it with the fluctuation threshold, mark the effective disturbance point, extract the direction and combine it into a joint state vector, and pass it to the path modeling module;

[0062] The path modeling module is used to call the joint state vector, determine the directional consistency of adjacent nodes based on the topological relationship table, divide the active response chain into the abnormal response chain, use the dynamic time warping algorithm to align the response delay, output the directional propagation path set, and pass it to the spectrum matching module;

[0063] The spectrum matching module is used to extract the original data by calling the time stamp of the directional propagation path set, perform Fourier transform to extract the main frequency band, compare the energy to mark the abnormal frequency band, output the frequency response correlation path, and transmit it to the fault location module;

[0064] The fault location module is used to retrieve the device mapping table to obtain the part number based on the starting node of the frequency response association path, verify the real-time parameter deviation of the valve opening and pump speed, and output the fault part number if it exceeds the tolerance threshold.

[0065] The real-time disturbance sequence specifically includes flow rate offset, pressure offset, and temperature offset. The joint state vector includes the effective disturbance coordinate set, the disturbance direction characteristic matrix, and the normalized fluctuation amplitude. The directional propagation path set specifically refers to the active response topology chain, the abnormal response topology chain, and the delay compensation coefficient. The frequency response association path includes the main frequency band amplitude spectrum, the abnormal frequency band energy distribution, and the phase alignment parameter. The fault component number specifically includes the valve opening deviation value, the pump speed deviation value, and the safe operation threshold.

[0066] Specifically, if Figure 2 As shown, the multi-source sensing module includes:

[0067] The multi-source sensor acquisition submodule collects flow pulse signals, pressure voltage signals, and temperature resistance signals, performs analog-to-digital conversion and fixed-time interception, performs multi-sensor timing alignment, and generates a multi-source time window data set;

[0068] During signal acquisition, sampling points are first deployed. For example, three flow rate sensors (numbered F1 to F3) are placed on the HVAC system's main circulation pipeline, located at the beginning, middle, and end of the pipeline. Two pressure sensors (numbered P1 and P2) are also placed at the compressor outlet and return inlet, and two temperature sensors (numbered T1 and T2) are placed at the air heating chamber inlet and outlet. All sensors are connected to the collector module. When collecting flow rate pulse signals, sampling is performed at a frequency of 1000Hz, obtaining 1000 pulse data points per second. For example, at a certain moment, the pulse count for F1 is 8730 within 10 seconds, and the calculated average frequency is 873Hz. Combined with the probe characteristics (each 10Hz represents 0.15m / s), the flow rate for F1 is 13.095m / s. The same calculation is performed for the remaining sensors. After the pressure signal is acquired and converted to voltage, it undergoes analog-to-digital conversion using a 12-bit ADC. For example, if P1 corresponds to a voltage of 2.35V and P2 to 1.80V, then within the 0-5V voltage range corresponding to a pressure of 0-1.6MPa, P1 is 0.752MPa and P2 is 0.576MPa. After the data is formatted uniformly, a complete sensor sample is captured every 5-second window, generating a set of raw data frames. Next, during the alignment process, the sensor clock offset is taken into account. The acquisition device timestamp is calibrated using the system clock reference signal, and the data frames are arranged according to the system's unified time sequence, so that the sample values ​​are aligned with a time interval of t=0s, t=5s, t=10s, and so on. If data synchronization is lost due to a delayed acquisition at T1 at the 6th second, the data at that time is interpolated between the preceding and following points. Ultimately, each 5-second window data set corresponds to nine parameter values: three flow rate values, two pressure values, two temperature values, and two timestamp error compensation parameters. The above process is sampled continuously to generate multiple time window data frames. For example, 12 sets of time window data sets are generated within a 60-second duration, where each set of data is as follows:

[0069] Table 1: Example table of multi-source time window dataset

[0070]

[0071] As shown in Table 1, parameter data collected every 5 seconds is integrated into a unified time window frame. Sensor signals undergo unified timestamp compensation, frequency or voltage conversion, and fixed sampling frequency analog-to-digital conversion before being combined to form a multi-source time series data set with real-time operational significance. This result forms a multi-dimensional, multi-time point mapping sequence of thermal and flow physical quantities that can be used for subsequent dynamic offset calculation operations.

[0072] The dynamic offset calculation submodule calls the multi-source time window data set, calculates the absolute value of the flow velocity offset at the beginning and end, analyzes the pressure window standard deviation, extracts the temperature linear trend slope, and sums the three parameter offsets weighted by the range ratio to generate a dynamic offset coefficient set;

[0073] Call the multi-source time window data set. First, compare the current sampling values ​​of the flow velocities F1, F2, and F3 in each time window data set with the sampling values ​​in the previous time window one by one, calculate the head and tail offsets of the flow velocity sensor, and determine the offset amplitude by the absolute value of the difference between the two moments. For example, at t=5s, the flow velocity of F1 is 13.05m / s, and at t=0s, it is 12.90m / s. Its offset is |13.05-12.90|=0.15m / s. The offsets of F2 and F3 are calculated to be |13.12-13.10|=0.02m / s and |13.18-13.20|=0.02m / s respectively. Then, the three offsets are calculated. The flow rate sensors are weighted separately, and the weights of flow rate sensors are set to F1=0.5, F2=0.3, and F3=0.2. The three-way weighted offset value is (0.5×0.15)+(0.3×0.02)+(0.2×0.02)=0.075+0.006+0.004=0.085m / s. The absolute value of the flow rate offset is calculated. Then, the standard deviation of the pressure parameters P1 and P2 is calculated for the sampling points within 5 seconds in the current time window. The sampling values ​​of P1 in the current window are 0.750, 0.752, and 0.755MPa, and the sampling values ​​of P2 are 0.578, 0.576, and 0.574MPa. The standard deviation is calculated by the formula: .

[0074] First calculate the mean: the mean of P1 is (0.750+0.752+0.755) / 3=0.752MPa, the mean of P2 is (0.578+0.576+0.574) / 3=0.576MPa, and then calculate the standard deviation of P1:

[0075] MPa; the P2 standard deviation is:

[0076] MPa;

[0077] The two sets of standard deviations are weighted as P1=0.6 and P2=0.4, and the weighted standard deviation is (0.6×0.00208)+(0.4×0.00163)=0.001248+0.000652=0.0019MPa. The pressure window standard deviation calculation is completed. Then, for temperatures T1 and T2, the temperature values ​​within the window are analyzed and the linear trend slope is fitted. The T1 temperature series is 23.5, 23.6, and 23.7℃, and the corresponding time series is 0, 5, and 10 seconds. The slope formula is:

[0078] ;

[0079] in 、 、 、 、 , substitute into the calculation:

[0080] ℃ / s;

[0081] The slope of T2 is calculated in sequence to obtain 0.03℃ / s, and the weights of T1=0.55 and T2=0.45 are assigned. The weighted temperature slope is (0.55×0.02)+(0.45×0.03)=0.011+0.0135=0.0245℃ / s. Finally, the flow rate, pressure, and temperature results are normalized and weighted and summed according to the range ratio. Assuming that the maximum range offsets are flow rate 1.0m / s, pressure 0.01MPa, and temperature 0.1°C / s, the corresponding normalized values ​​are 0.085 / 1.0=0.085, 0.0019 / 0.01=0.19, and 0.0245 / 0.1=0.245, respectively. According to the weights of flow rate 0.4, pressure 0.3, and temperature 0.3, the weighted synthesis is , and the dynamic offset coefficient of this time window is 0.1645.

[0082] The perturbation sequence generation submodule fits the offset coefficient change curve based on the dynamic offset coefficient set, calculates the change rate using a sliding window, and compares it with the original percentile threshold. When the threshold is exceeded three times in a row, a real-time perturbation sequence is generated;

[0083] Based on the dynamic offset coefficient set, the dynamic offset coefficients continuously calculated under the aforementioned multi-window are first arranged into a sequence. For example, the offset coefficients corresponding to the first 12 groups of windows are 0.158, 0.162, 0.1645, 0.167, 0.169, 0.172, 0.175, 0.179, 0.180, 0.183, 0.185, and 0.188, respectively. Then, the sliding window method is used to extract 3 consecutive groups of coefficient segments for change rate calculation. The step size is 1 group, and the calculation method is change rate = (current value - previous value) / previous value. For example, the change rates of the 2nd to 4th groups are: (0.162-0.158) / 0.158≈0.0253, (0.1645-0.162) / 0.162≈0.0154, (0.167-0.1645) / 0.1645≈0.0152, and then each group is taken. The average of the three change rates is taken as the average change rate of the paragraph, which is (0.0253+0.0154+0.0152) / 3≈0.01863. The paragraph change rate sequence is then sorted into 0.01863, 0.0187, 0.0178, 0.0195, 0.0189, 0.0194, 0.0183, 0.0188, 0.0190, and 0.0185, respectively. The original percentile threshold is set to 0.0190, that is, when the average change rate is greater than the threshold, it is determined to be exceeded. Subsequently, the average change rate of each group of paragraphs is compared one by one. The 4th group (0.0195), the 6th group (0.0194), and the 9th group (0.0190) all reach or exceed the threshold. If the threshold is exceeded three times in a row, it is determined that a disturbance event has occurred. In this example, the 4th, 5th, and 6th groups exceed the threshold continuously, so a real-time disturbance sequence is generated.

[0084] Specifically, if Figure 2 As shown, the disturbance feature module includes:

[0085] The disturbance intensity quantification submodule reads the absolute disturbance sequence, performs intensity marking processing, compares the preset dynamic fluctuation threshold point by point, calculates the percentage of points exceeding the threshold, and generates the disturbance intensity coefficient;

[0086] The fluctuation threshold setting method is based on the mean of the original parameter data plus or minus two standard deviations. It is obtained by analyzing real-time operating data and experimentally verifying 10 HVAC equipment, covering three typical models. The sampling frequency is 100Hz and the confidence interval is set to 95%.

[0087] When the disturbance intensity quantification submodule reads the disturbance absolute sequence, it first extracts the corresponding point pressure data according to the monitoring time series of the equipment and constructs the disturbance absolute value sequence. During the execution process, the sampling data 100 times per second is preprocessed, including data denoising, missing interpolation, unit conversion and other processing steps. The specific method is to apply a mean filter to the 100 pressure values ​​per second, remove more than 5% of the peak abnormal points and then process them. Taking Unit 1 in the HVAC system as an example, its monitoring pressure sequence is Calculate the absolute perturbation sequence, that is, perform the absolute difference operation on each two adjacent sampling points to form , then a point-by-point comparison is performed based on the dynamic fluctuation threshold. The threshold is the mean plus or minus two times the standard deviation. The statistical method is to calculate the mean and standard deviation of all data points in the sequence, and set the confidence interval to 95%. Taking the data of 1 of 10 devices as an example, assuming that the sample mean is 0.42Pa and the standard deviation is 0.15Pa, the upper threshold limit is , the lower limit is , since it is an absolute value sequence, the upper limit of 0.72Pa is used as the judgment standard, and each value in the sequence is traversed in order and the number of points greater than the threshold is recorded, and then the ratio is calculated with the total number of points in the sequence. Suppose the sequence has 5000 points, and the number of points exceeding the threshold is 480 points, then the ratio is , and use this proportion as the disturbance intensity coefficient , the generated value is 0.096, indicating that the disturbance intensity in this data window is in the medium to low range. The reference range is divided into: , for low disturbance, , which is a medium disturbance, and greater than 0.15, which is a high disturbance level. The final disturbance intensity coefficient is 0.096.

[0088] The effective disturbance identification submodule calculates the dynamic threshold adjustment factor based on the disturbance intensity coefficient, marks the coordinates of the points exceeding the threshold, uses the sign function to extract the positive and negative polarity features, constructs a polarity-coordinate mapping table with a timestamp, and generates a valid disturbance coordinate set;

[0089] The effective disturbance identification submodule calls the aforementioned disturbance intensity coefficient Then, calculate the dynamic threshold adjustment factor , using the formula:

[0090] ;

[0091] in, Represents the percentage adjustment factor of the dynamic fluctuation threshold at time t, which is a dimensionless parameter. Represents the weight coefficient of the disturbance intensity coefficient, which is a dimensionless parameter. Represents the perturbation intensity coefficient generated by the previous step, which is a dimensionless parameter. Represents the adjustment factor of the gradient change at the current moment, in s / Pa. Represents the discrete gradient value of the disturbance sequence at time t, in Pa / s, Represents the original data window length, Represents the attenuation weight of the i-th original window, in units of 1 / s, Represents the nonlinear activation coefficient, which is a dimensionless parameter. Represents the integrated value of the disturbance energy in the i-th window, in Pa 2 ·s.

[0092] First, get the discrete gradient value at the current time t , specifically the difference between the tth point and the previous point in the sequence, in Pa / s, with a sampling frequency of 100 Hz, then , if the pressures at the 2001st and 2000th points are 204.2Pa and 203.7Pa respectively, then ,set up 、 、 , the integrated value of disturbance energy is the integral of the square of the disturbance value in the i-th data window, and the samples of three data windows are , and its corresponding weight is set to , put the parameters into the formula and calculate as follows:

[0093] ;

[0094] The results show that the dynamic threshold needs to adjust the initial fluctuation threshold by 357.4%, and then re-traverse the threshold points with the adjusted threshold, record their coordinates, and determine the corresponding The sign value is constructed in the positive and negative directions (positive is +1, negative is -1), and the polarity-coordinate mapping table is generated in conjunction with the timestamp. For example, the pressure change at point 1201 is +0.8Pa, the sign is +1, and the timestamp is "2025-06-12-14:23:21.120". The mapping table record is (+1, 1201, 2025-06-12-14:23:21.120), and finally a polarity perturbation coordinate set consisting of super-threshold effective perturbation points is generated.

[0095] The state vector synthesis submodule calls the polarity characteristics of the effective perturbation coordinate set, normalizes the amplitude of the positive and negative polarities using the range method, and concatenates the normalized amplitude with the original coordinates in time sequence to generate a joint state vector.

[0096] The state vector synthesis submodule extracts the polarity value of each record based on the polarity disturbance coordinate set, and classifies it into two sets according to positive and negative classification. The maximum and minimum amplitudes are taken to form the range interval, and the amplitude normalization operation is performed. For example, the maximum disturbance amplitude in the positive set is 0.85Pa and the minimum is 0.35Pa, and the corresponding interval is , the normalized calculation formula is: , C is the current amplitude. If the current amplitude is 0.55Pa, the normalized value is , the negative electrode set is processed in the same way. The normalized amplitude is concatenated with the original coordinates and arranged in ascending time order to form a joint state vector set. The joint vector example is: .

[0097] Table 2: Polarity disturbance amplitude normalization results

[0098]

[0099] As shown in Table 2, the normalized processing results of four typical disturbance points are listed. Combined with the amplitude information and polarity mapping relationship of the time point, it can be further used for subsequent operations such as state space modeling.

[0100] Specifically, if Figure 2 As shown, the path modeling module includes:

[0101] The directional consistency judgment submodule calls the connection relationship between the node displacement parameters in the joint state vector and the topological relationship table, extracts the displacement components of the node on the X-axis and Y-axis to construct a two-dimensional displacement vector, calculates the cosine value of the displacement direction angle between adjacent nodes based on the vector dot product formula, compares the cosine value with the directional consistency threshold item by item, and generates the node offset coefficient;

[0102] The node offset coefficient indicates the intensity of the node response trend change and serves as a reference for chain division;

[0103] The direction consistency judgment submodule calls the connection relationship between the node displacement parameters in the joint state vector and the topological relationship table. It is necessary to first extract the displacement components of each pair of adjacent nodes on the X-axis and Y-axis to construct a two-dimensional displacement vector. For example, the displacement of node A is , node B is , and their two-dimensional displacement vectors are , then calculate the dot product of the two vectors, that is , and calculate the modulus at the same time. The modulus of node A is , the module length of node B is , based on which the cosine of the angle between the two vectors is , and then call the threshold judgment rule to compare the cosine value with the direction consistency threshold. The direction consistency threshold can be preset to 0.9, that is, when the angle cosine value ≥ 0.9, it is determined that the direction is consistent. If the cosine value is < 0.9, it is considered that there is a direction offset, and then the offset coefficient is calculated. Here, the offset coefficient is set to ,Right now The higher the offset coefficient, the greater the difference in displacement direction. According to this type of calculation logic, the adjacent nodes are iteratively processed to obtain the directional offset coefficient set of the complete node pair and record it, thereby providing data support for the subsequent response chain division. The node offset coefficient can be compared with the set multi-level interval threshold, such as 0-0.1 for the first level (direction convergence), 0.1-0.3 for the second level (slight offset), 0.3-0.6 for the third level (obvious offset), and greater than 0.6 for the fourth level (strong offset). The calculation nodes are classified and statistically formed into a directional offset level set.

[0104] Table 3: Calculation table of node pair direction cosine value and offset coefficient

[0105]

[0106] Table 3 lists the calculated cosine values ​​of the direction angles and the offset coefficients for four typical node pairs. It can be seen that the node pairs AB have completely consistent directions, with an offset coefficient of 0.000, while the node pairs GH have completely opposite directions, with an offset coefficient of 2.000, which is a serious anomaly. This table will be used in response chain segmentation as an important parameter for node behavior classification.

[0107] The response chain division submodule establishes a node status classification matrix based on the node offset coefficient, divides the coefficient distribution into quartile intervals, identifies nodes in the upper and lower quartiles and classifies them into active chains and abnormal chains, counts the spatial coordinates of nodes in the chain, and generates a chain identification matrix;

[0108] The node offset coefficient represents the intensity of the node response trend change. It is necessary to establish a node state classification matrix based on the obtained node offset coefficient sequence and the statistical distribution law. In this process, the node offset coefficient values ​​​​need to be integrated and summarized, and the offset coefficients of the adjacent edges of each node are averaged to form the comprehensive offset index of the node. For example, the offset coefficients of a node A and its three adjacent nodes B, C, and D are 0.068, 0.224, and 0.432 respectively. The comprehensive offset index of node A is ,Then the node comprehensive offset index is formed into a one-dimensional vector sequence and arranged in ascending order. Then the quartiles Q1 (25% quantile), Q2 (median), and Q3 (75% quantile) of the sequence are calculated to divide the offset index into four intervals. Assuming that the offset index sequence collected from a project measurement is [0.000, 0.068, 0.102, 0.148, 0.192, 0.241, 0.302, 0.428, 0.589, 0.776], then Q1 is 0.102 and Q3 is 0.428 through quantile calculation. Nodes below Q1, such as 0.000 and 0.068, are classified as abnormal chains, and nodes above Q3, such as 0.589 and 0.776 are classified as active links, and the nodes in the middle are not classified for the time being. Next, the spatial coordinate information of each node identified in the chain needs to be counted. The physical coordinate position corresponding to the node number can be read from the initial sensor point table. For example, the coordinates of node A are (5.2, 8.3) and those of node B are (6.1, 7.4). The numbers and coordinates of the nodes in the chain are combined into a chain identification matrix. Each row represents a triple consisting of the number, offset value, and coordinate information of a node in the chain. During the execution process, the identified chain node set needs to be processed cyclically to generate the active chain identification matrix and the abnormal chain identification matrix respectively, providing the chain infrastructure data for the subsequent response chain time series processing module.

[0109] The delay alignment submodule uses the chain identification matrix to extract the time series of the active response chain and the abnormal response chain, constructs the cost matrix of the dynamic time warping algorithm, recursively searches the accumulated path alignment waveform, uses cubic spline interpolation to fill in the gaps, and generates a set of directed propagation paths;

[0110] The delay alignment submodule calls the chain identification matrix generated above and needs to extract the response time series corresponding to the active response chain and the abnormal response chain respectively. The time series data can be extracted through the system monitoring data log. For example, the response time series of node A is , node B is , the two sequences are used as the input data of the dynamic time warping process to construct the cost matrix. The element value of the cost matrix is ​​the absolute value of the corresponding time point difference. For example, the first row and first column element is , according to the rule, a 4×4 dimensional matrix is ​​constructed and filled in sequence, and then the cumulative path is recursively searched from the upper left corner to the lower right corner. At each step, the minimum cost value among the three candidate paths is selected and accumulated. Finally, the path with the minimum cost is obtained as the alignment path of the two sequences, and the paired time point sequence is obtained. For example, the alignment result is If there is a vacant position in the alignment path, that is, a node response fails to match the paired point, the cubic spline interpolation function is called to interpolate and complete the gap, for example, at the time point in the active link , there is no matching value, then construct a spline curve in the corresponding interval in the abnormal chain and interpolate to generate Finally, the alignment path completion is completed to form a paired time series path set. The path set format is a triple of node pair number, active link response time point, and abnormal link pairing time point. Each set of paths represents a directional propagation path and is recorded in the path set for analyzing fault propagation rules.

[0111] Specifically, if Figure 2 As shown, the spectrum matching module includes:

[0112] The time series data acquisition submodule obtains the time stamp of the directional propagation path set, calls the original signal data of the corresponding time interval, detects the sampling rate matching, interpolates and compensates for the offset time series segment, and generates the original signal time series set;

[0113] In the time series data acquisition submodule, when obtaining the timestamp of the directional propagation path set, the system calls the recorded device operation log and the path propagation time information recorded by the sensor network, and associates the path number with the timestamp. For example, in the 8 paths arranged between the air-conditioning host and the terminal air outlet, the path propagation records between 10:02:15 and 10:07:15 on May 15, 2025 are recorded respectively, forming a mapping sequence of numbered paths P1~P8 and time T1~Tn. In the process of calling the original signal data of the corresponding time interval, the temperature, humidity, wind speed, and pressure signal sampling values ​​stored in the industrial edge gateway or SCADA system are read to extract the original sampling data between time T1 and Tn in the form of an array. The interval between each sampling point does not exceed 0.1 second. For example, at time T1=10:03:05, the signal sequence collected by device P1 is When checking the sampling rate matching, the path sampling frequency is compared with the system standard frequency. For example, the real-time sampling rate of path P1 is 9.8Hz, while the system set sampling rate is 10Hz. At this time, the calculated sampling deviation is 0.2Hz, and it is marked as an object that needs interpolation processing. When interpolating and compensating the offset timing segment, the linear interpolation method between two adjacent points is used. That is, if there is a missing or deviation point between 10.03 seconds and 10.13 seconds, the defect value is linearly calculated using the values ​​of the two points. For example ( ), the interpolation value is In the process of generating the original signal time series set, the signal values ​​after path completion are rearranged according to the unified time axis to form P1: P2: Table 4 lists the signal sampling values ​​and interpolation processing of the two paths at different time points.

[0114] Table 4: Sensing path sampling data table

[0115]

[0116] As shown in Table 4, there is a missing signal at 10.13 seconds in the P1 path. The system automatically interpolates the adjacent values ​​on both sides to complete the time series signal content, thereby ensuring the continuity of subsequent frequency domain analysis.

[0117] The frequency domain conversion submodule performs Fourier transform based on the original signal time series set, calculates the frequency band spectrum amplitude, extracts the main frequency band parameters, and simultaneously calculates the energy distribution characteristics of the original normal working condition to establish the main frequency band energy feature set;

[0118] When the frequency domain conversion submodule performs Fourier transform based on the original signal time series set, the P1 path temperature signal sequence collected by the SCADA system , (sampling rate 10Hz, time window 256 points) discrete Fourier transform, frequency resolution , the 5th frequency point corresponds to the frequency , plural result , calculate the energy value , weight factor According to the vibration energy transfer experiment, it is set to {0.5, 1.0, 0.5} (comparison test shows that the sensitivity is improved by 12%), and the phase difference Calculated by the phase angle difference of adjacent frequency points: , ,have to , after normalization , power change rate Calculated from the energy difference: , , , response time (From the duct system test report), , the mean baseline energy , when calculating the frequency band spectrum amplitude, bring in the following calculation formula:

[0119] ;

[0120] in, represents the spectrum amplitude of the kth frequency band, Represents the energy value of the k+mth frequency point, which comes from the Fourier transform result of the original signal time series set, and the unit is J. Representative The average value of the reference energy of three adjacent frequency points in the frequency band, in J. Represents the adjacent frequency weight factor , Represents the energy value of the k+nth frequency point after Fourier transform of the original signal time series set, the unit is J, Representative Normalized vibration phase difference of the frequency band, Representative Frequency band power change rate, unit is J / s, System characteristic response time, in seconds.

[0121] ;

[0122] The results show that the spectrum amplitude of the 5th frequency band is 0.1581, which is lower than the set threshold of 0.2 in the spectrum amplitude distribution range, and is classified as an area with insignificant energy fluctuations. The benefit of the formula is that it effectively extracts the frequency band where structural changes appear by jointly calculating the energy fluctuations and vibration phase changes of adjacent frequency points and introducing a normalization factor for the total energy.

[0123] The frequency band anomaly detection submodule calls the main frequency band energy feature set, separates the current energy spectrum from the baseline energy distribution, calculates the frequency band energy difference, marks the abnormal frequency band three-dimensional parameters, integrates the abnormal frequency band mark set, and generates the frequency response correlation path;

[0124] The frequency band anomaly detection submodule calls the main frequency band energy feature set, the main frequency band of the current path P1~P8 The value is compared with the statistical energy value of the corresponding frequency band in the benchmark energy distribution, and the deviation points are judged, marked and classified according to the preset statistical threshold and interval settings, and then the marked frequency bands are integrated to form an abnormal frequency band set, thereby generating a frequency response correlation path. The specific execution process is as follows:

[0125] First, read the energy mean and standard deviation of each path P1~P8 in the main frequency band in the benchmark energy distribution feature set, for example, the energy mean of frequency band k=5 in the benchmark set is , standard deviation ; At the same time, obtain the P1 path frequency band k=5 in the current monitoring data and calculate , but its corresponding energy value needs to be adjusted back to the energy dimension, and the current energy can be inferred: Substitute the known total energy ,then , that is, the current energy is 2.0J, which is far lower than the benchmark mean of 16.0J; then, the normal energy range is constructed based on the benchmark mean and standard deviation , this interval is used as the "normal" standard. If the current energy is lower than 13.0J, it is judged as "low abnormality", and if it is higher than 19.0J, it is judged as "high abnormality". In this example, 2.0J is lower than 13.0J and is marked as low abnormality. Subsequently, the comparison and judgment process is executed for each path and each main frequency band in turn, and the marking results are stored in the form of frequency band energy matrix dimensions, where the path and frequency band constitute a two-dimensional index, and the value set is {"normal", "low abnormality", "high abnormality"}.

[0126] Secondly, a threshold setting mechanism is used to aggregate abnormal frequency band marks. According to statistical rules, if a single path has two or more frequency bands judged as "low abnormality" or "high abnormality" in three consecutive main frequency bands, the path is marked as an "abnormal path"; for example, P1 obtains the current energy in frequency bands k=4, 5, and 6 respectively. (Benchmark , [ ]=[13,19],[ ]=[14±3]=[11,17]), among which 2.5<9, 2.0<13, and 3.0<11, all three frequency bands are "low abnormality", thus satisfying the "two or more" condition. Path P1 is marked as an abnormal path and written into the abnormal path set.

[0127] Afterwards, the path and the corresponding abnormal frequency band set are integrated to generate a "frequency response associated path" structure, which includes the path number, abnormal frequency band index and corresponding abnormal type. For example, the output structure is as follows: When generating the structure, the abnormal frequency band numbers are sorted in ascending order, and isolated points with frequency band intervals greater than 2 are removed to ensure that the associated path frequency bands are structural connection areas. The specific processing actions are: Read the path P1 marked frequency band set , the difference between adjacent frequency bands is 1 less than or equal to 2, and all are retained; if the P2 set is {3, 7}, and the difference is 4 greater than 2, then it is split into two records .

[0128] Finally, the system sorts the abnormal path set in ascending order by path number for subsequent visualization or fault location. After the path numbers are sorted, the system outputs the final frequency response associated path array and writes the array to the database or exports it as a JSON format message for subsequent use.

[0129] Specifically, if Figure 2 As shown, the fault location module includes:

[0130] The component association retrieval submodule obtains the starting node number of the frequency response association path, matches the relationship between the node and the component number, reads the unique identifier corresponding to the current node, and generates the component association number;

[0131] The component association retrieval submodule first receives the starting node number of the frequency response association path from the fault alarm module, for example, number 12. On this basis, it performs the corresponding mapping operation between the node number and the equipment part number, retrieves the component mapping relationship table established in the system database, and confirms whether the structural module to which the numbered node 12 belongs is an executive component or a sensing component. If it is confirmed to be a valve body, its associated unique identifier "V12" is extracted. In this process, the one-to-one correspondence between the three database fields "node number-component number-unique identifier" is read, and the corresponding record is retrieved through hash matching or array index comparison method. In a real scenario, for example, the numbered node 12 in the HVAC system pipeline is a hot and cold mixing control valve, and the control signal of the valve comes from the output signal of the PID controller. Its physical number is CV01, and the corresponding logical number is set to V12. V12 is further used as the component retrieval number. The connection nodes and original states related to V12 are extended and read to construct a logical linked list of the initial components of the frequency response association path. The related valve numbers under the same path are then filtered based on this linked list. For example, if V13 and V14 exist in the path, the component set {V12, V13, V14} of the current frequency response association path is generated in the form of a list. In this execution, the confirmation of the frequency response association path is determined by the abnormal value of the trigger frequency of the system fault sensor. If the collected frequency value at the starting point of the path exceeds its set tolerance by more than 5%, it is judged as an abnormal start node. Then, the component numbers of all nodes involved in the path are merged, and finally a component association number set is generated for subsequent calls.

[0132] Table 5: Example table of component association retrieval

[0133]

[0134] As shown in Table 5, the mapping process between node numbers and unique identifiers is established in a predefined structural relational database. When node number 12 is identified as an abnormal frequency response source, the system uses this node as the core and traces back to the execution component corresponding to its control logic number Ctrl-03. Using the component's unique identifier V12 as the association item, the system gradually calls the logical numbers of the remaining components physically connected to it, such as V13 and V14. The comparison process performed in this identification process is as follows: traverse the nodes on the current path, determine each one's direct connection to the starting node, and if so, add its logical identifier to the association set, ultimately constructing a complete component chain. For example, in a commercial building HVAC system, if the initial node frequency fluctuation exceeds 5% of the allowable fluctuation range (i.e., the starting frequency is 45 Hz and the rated frequency is 42 Hz), the node is marked as an abnormal source node by the system. From this, the correspondence between the starting number 12 and the unique identifier V12 is derived, and a component association number set {V12, V13, V14} with V12 as the header is constructed for subsequent dynamic parameter acquisition and call.

[0135] The real-time parameter acquisition submodule reads the valve sensor pulse voltage signal and the pump body frequency current signal based on the component association number, converts them into standard opening percentage and standard speed range, and generates dual-channel dynamic parameters;

[0136] After receiving the component association number set {V12, V13, V14}, the real-time parameter acquisition submodule locates the device type and installation position corresponding to the number to read the sensor data stream bound to it. Among them, V12 corresponds to the hot and cold mixing control valve, which is installed with an angular stroke electric valve opening sensor. The sensor output is a 0~10V pulse voltage signal, and the sampling period is set to 0.5 seconds. By continuously reading a total of 20 groups of pulse voltage signal data within 10 seconds, the current average pulse voltage of the valve is calculated to be 6.4V. Then, according to the device setting, the opening and voltage linear mapping is calculated. The relationship between V and V is that 0V represents 0% opening, 10V represents 100% opening, and the current standard opening percentage is 64%. At the same time, the same operation is performed on V13 and V14, and the voltage values ​​are 4.2V and 8.1V respectively, corresponding to the opening percentages of 42% and 81%. For the pump body components, for example, there is a circulation pump numbered P01 at the end of the path. Its sensor is a current frequency acquisition module based on the Hall effect, and the output is a 0~20mA current signal. 20 groups of current values ​​within 10 seconds are read through the PLC sampling interface, and the current pump body operating frequency is obtained after converting them into frequency. The frequency is 38.7Hz, and the standard setting frequency range is 30Hz to 45Hz. The current frequency is in the middle of operation. The standard speed range is calculated as: (38.7-30) / (45-30)=58%. The reading operation obtains its sensor interface channel by matching the device ID, and then sequentially retrieves the original electrical signal value of the device from the real-time data acquisition terminal. During signal reading, each sensor has a unique communication address. The PLC or DCS system reads the signal value through the ModbusRTU protocol polling. After each reading, the result is sent to the submodule cache through the scheduling thread. Store the queue and establish a mapping dictionary between the device number and the sensor output in the memory. The voltage and current value conversions are based on the set linear conversion coefficient. For example, the valve opening conversion coefficient is 10%, that is, every 1V represents 10% opening, and the pump body frequency conversion coefficient is 0.75Hz / mA, that is, every 1mA represents 0.75Hz frequency. Finally, a dual-channel dynamic parameter group is formed with a format such as {V12: opening 64%, V13: opening 42%, V14: opening 81%, P01: frequency 38.7Hz (accounting for 58%)} for the next step of deviation calculation.

[0137] The deviation threshold determination submodule calls dual-channel dynamic parameters, calculates the opening difference and speed change rate, synchronously compares the system tolerance threshold, and outputs the fault component number;

[0138] The tolerance threshold fluctuates 5% based on the upper limit of the parameter rating. The error rate is less than ±0.8% after 40 dynamic loadings in the laboratory and 3 days of continuous operation in the production environment.

[0139] The deviation threshold determination submodule calls the above dynamic parameter group and performs a calculation on the deviation between the current opening and the rated opening under the original stable working condition for each valve. For example, the rated opening of V12 is 70%, the current opening is 64%, and the calculated opening difference is 70%-64%=6%. For the pump body P01, the speed change rate is calculated. The original frequency setting value is 42Hz, and the current value is 38.7Hz. The change rate is |(42-38.7) / 42|=7.86%. At the same time, each calculation result is compared with the set tolerance threshold. The tolerance threshold has been tested by 40 dynamic loading tests in the laboratory. After three days of continuous industrial operation verification, the final deviation limit was set to 5%. During the calculation process, the threshold was exceeded by comparing each item one by one. The specific actions are: if the opening difference or change rate is greater than 5%, the component is marked as a faulty component, and the subsequent component numbers are traversed, and the difference and ratio between them and the rated value are calculated one by one. For example, the rated opening of V13 is 45%, and the current opening is 42%. The difference is 3%, which is less than the threshold, and no fault is marked. The rated opening of V14 is 78%, and the current opening is 81%. The difference is 3%, and no fault is marked either. Therefore, only V12 and P01 are finally identified as faulty by the system. The faulty component number is generated based on this, and the current fault component number set {V12, P01} is generated. During the execution process, the "judgment" operation is based on whether it is greater than 5% as the only condition, and the "calculation" operation is based on the numerical difference analysis between the real-time sampling value and the rated value and outputs the specific percentage. The default comparison value inside the system retains two decimal places. In the data structure, a dictionary structure is used to record each component number and its difference percentage and marking status, such as {V12: 6%, status: fault}, {V13: 3%, status: normal}. Fuzzy language is not used in the judgment logic, but a set numerical range is used. For example, if the threshold is set to 5%, the judgment criteria are 0-5% as the normal range and >5% as the fault range. This range division is based on the stability analysis of the original data. The specific setting process is: based on the rated value of each component, the floating upper and lower limits are ±5%. That is, if V12 is rated at 70%, the threshold range is [66.5%, 73.5%]. As long as the real-time data falls outside this range, it is a fault. Combined with this definition, through the calculation and comparison of the opening and frequency differences, the faulty component number set with the current abnormality is finally identified as {V12, P01}.

[0140] See also Figure 3 The HVAC system fault location method based on big data is executed based on the above-mentioned HVAC system fault location system based on big data, and includes the following steps:

[0141] S1: Flow rate, pressure, and temperature values ​​are collected through a multi-channel sensor group. The sliding window algorithm is used to segment the continuous time series data into equal lengths. The offset between the parameters in the window and the reference value is calculated to generate a real-time disturbance sequence.

[0142] S2: Call the real-time disturbance sequence, input the data points into the absolute value operation function for amplitude conversion, filter the effective disturbance points through the preset fluctuation threshold, extract the change direction of adjacent disturbance points, and combine the joint state vector;

[0143] S3: Based on the node topology of the joint state vector, a directional consistency determination algorithm is used to divide the response chain types. The active response chain and the abnormal response chain are input into the dynamic time warping algorithm for delay alignment, and a set of directional propagation paths is output.

[0144] S4: Call the timestamp index of the directional propagation path set, extract the corresponding original vibration data, calculate the spectrum distribution through fast Fourier transform, compare the energy ratio threshold of the standard frequency band, mark the abnormal frequency band, and generate the frequency response correlation path;

[0145] S5: According to the starting node number of the frequency response association path, the device topology database is retrieved to obtain the corresponding component code, and the valve opening setting value and the pump speed feedback value are compared in real time. When the deviation exceeds the preset tolerance, the fault component number is output.

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A HVAC system fault location system based on big data, characterized in that: The system comprises: The multi-source sensing module is used to collect flow rate, pressure, and temperature through a sensor group, split the data into time windows, calculate parameter offsets, generate real-time disturbance sequences, and pass them to the disturbance feature module; A disturbance feature module is used to calculate the absolute value of the real-time disturbance sequence, compare it with the fluctuation threshold, mark the effective disturbance point, extract the direction and combine it into a joint state vector, and pass it to the path modeling module; A path modeling module is used to call the joint state vector, determine the directional consistency of adjacent nodes based on the topological relationship table, divide the active response chain into an abnormal response chain, align the response delay using a dynamic time warping algorithm, output a directional propagation path set, and pass it to the spectrum matching module; A spectrum matching module is used to extract raw data by calling the timestamp of the directional propagation path set, perform Fourier transform to extract the main frequency band, compare the energy to mark the abnormal frequency band, output the frequency response correlation path, and transmit it to the fault location module; a fault location module, configured to retrieve a device mapping table to obtain a component number based on a starting node of the frequency response association path, verify a real-time parameter deviation of the valve opening and the pump speed, and output a fault component number if the deviation exceeds a tolerance threshold; The disturbance feature module includes: The disturbance intensity quantification submodule reads the real-time disturbance sequence, performs intensity marking processing, compares it point by point with a preset dynamic fluctuation threshold, calculates the percentage of points exceeding the threshold, and generates a disturbance intensity coefficient; The effective disturbance identification submodule calculates the dynamic threshold adjustment factor based on the disturbance intensity coefficient, marks the coordinates of the points exceeding the threshold, extracts the positive and negative polarity features using a sign function, constructs a polarity-coordinate mapping table with a timestamp, generates an effective disturbance coordinate set, and calculates the dynamic threshold adjustment factor using the following formula: ;in, Represents the percentage adjustment factor of the dynamic fluctuation threshold at time t, which is a dimensionless parameter. Represents the weight coefficient of the disturbance intensity coefficient, which is a dimensionless parameter. Represents the perturbation intensity coefficient generated by the previous step, which is a dimensionless parameter. Represents the adjustment factor of the gradient change at the current moment, in s / Pa. Represents the discrete gradient value of the disturbance sequence at time t, in Pa / s, Represents the original data window length, Represents the attenuation weight of the i-th original window, in units of 1 / s, Represents the nonlinear activation coefficient, which is a dimensionless parameter. Represents the integrated value of the disturbance energy in the i-th window, in Pa 2 ·s; The state vector synthesis submodule calls the polarity characteristics of the effective perturbation coordinate set, normalizes the amplitude of the positive and negative polarities by the range method, and performs tensor splicing of the normalized amplitude and the original coordinates in chronological order to generate a joint state vector; The path modeling module includes: The directional consistency judgment submodule calls the connection relationship between the node displacement parameters in the joint state vector and the topological relationship table, extracts the displacement components of the node on the X and Y axes to construct a two-dimensional displacement vector, calculates the cosine value of the displacement direction angle between adjacent nodes based on the vector dot product formula, compares the cosine value with the directional consistency threshold item by item, and generates a node offset coefficient; the node offset coefficient represents the intensity of the node response trend change and serves as a reference for chain division; The response chain division submodule establishes a node status classification matrix based on the node offset coefficient, divides the coefficient distribution into quartile intervals, identifies nodes in the upper quartile and lower quartile and classifies them into active chains and abnormal chains, counts the spatial coordinates of the nodes in the chain, and generates a chain identification matrix; The delay alignment submodule calls the chain identification matrix to extract the time series corresponding to the active response chain and the abnormal response chain, which is used as the input data of the dynamic time warping algorithm to construct the cost matrix, recursively searches the cumulative path alignment waveform, uses cubic spline interpolation to fill in the gaps, and generates a set of directed propagation paths.

2. The HVAC system fault location system based on big data according to claim 1 is characterized by: The real-time disturbance sequence specifically includes flow rate offset, pressure offset, and temperature offset. The joint state vector includes a valid disturbance coordinate set, a disturbance direction characteristic matrix, and a normalized fluctuation amplitude. The directional propagation path set specifically includes an active response topology chain, an abnormal response topology chain, and a delay compensation coefficient. The frequency response association path includes a main frequency band amplitude spectrum, an abnormal frequency band energy distribution, and a phase alignment parameter.

3. The HVAC system fault location system based on big data according to claim 1 is characterized in that: The multi-source sensing module includes: The multi-source sensor acquisition submodule collects flow pulse signals, pressure voltage signals, and temperature resistance signals, performs analog-to-digital conversion and fixed-time interception, performs multi-sensor timing alignment, and generates a multi-source time window data set; The dynamic offset calculation submodule calls the multi-source time window data set, calculates the absolute value of the flow velocity offset at the beginning and end, analyzes the pressure window standard deviation, extracts the temperature linear trend slope, and sums the three parameter offsets weighted by the range ratio to generate a dynamic offset coefficient set; The disturbance sequence generation submodule fits the displacement coefficient change curve based on the dynamic displacement coefficient set, calculates the rate of change using a sliding window, compares the original percentile threshold, and generates a real-time disturbance sequence when the threshold is exceeded three times in a row.

4. The HVAC system fault location system based on big data according to claim 3 is characterized in that: The spectrum matching module includes: The time series data acquisition submodule obtains the timestamp of the directional propagation path set, calls the original signal data of the corresponding time interval, detects the sampling rate matching, performs interpolation compensation on the offset time series segment, and generates the original signal time series set; The frequency domain conversion submodule performs Fourier transform based on the original signal time series set, calculates the frequency band spectrum amplitude, extracts the main frequency band parameters, synchronously calculates the original normal working condition energy distribution characteristics, and establishes the main frequency band energy feature set; The frequency band anomaly detection submodule calls the main frequency band energy feature set, separates the current energy spectrum from the reference energy distribution, calculates the frequency band energy difference, marks the abnormal frequency band three-dimensional parameters, integrates the abnormal frequency band mark set, and generates a frequency response correlation path.

5. The HVAC system fault location system based on big data according to claim 4 is characterized in that: Calculate the spectrum amplitude of the frequency band using the formula: ； in, represents the spectrum amplitude of the kth frequency band, Represents the energy value of the k+mth frequency point, which comes from the Fourier transform result of the original signal time series set, and the unit is J. Representative The average value of the reference energy of three adjacent frequency points in the frequency band, in J. represents the adjacent frequency weight factor, , Represents the energy value of the k+nth frequency point after Fourier transform of the original signal time series set, the unit is J, Representative Normalized vibration phase difference of the frequency band, Representative Frequency band power change rate, unit is J / s, System characteristic response time, in seconds.

6. The HVAC system fault location system based on big data according to claim 5 is characterized in that: The fault location module includes: The component association retrieval submodule obtains the starting node number of the frequency response association path, matches the relationship between the node and the component number, reads the unique identifier corresponding to the current node, and generates a component association number; The real-time parameter acquisition submodule reads the valve sensor pulse voltage signal and the pump body frequency current signal based on the component association number, converts them into standard opening percentage and standard speed range, and generates dual-channel dynamic parameters; The deviation threshold determination submodule calls the dual-channel dynamic parameters, calculates the opening difference and the speed change rate, synchronously compares the system tolerance threshold, and outputs the fault component number.

7. A method for locating faults in a HVAC system based on big data, characterized in that: The method is used to implement the HVAC system fault location system based on big data according to any one of claims 1 to 6, comprising the following steps: S1: Flow rate, pressure, and temperature values ​​are collected through a multi-channel sensor group. The sliding window algorithm is used to segment the continuous time series data into equal lengths. The offset between the parameters in the window and the reference value is calculated to generate a real-time disturbance sequence. S2: calling the real-time disturbance sequence, inputting the data points into the absolute value operation function for amplitude conversion, screening the effective disturbance points through the preset fluctuation threshold, extracting the change direction of adjacent disturbance points, and combining the joint state vector; S3: Based on the node topology relationship of the joint state vector, a directional consistency determination algorithm is used to divide the response chain types, and the active response chain and the abnormal response chain are input into the dynamic time warping algorithm for delay alignment, and a directional propagation path set is output; S4: calling the timestamp index of the directed propagation path set, extracting the corresponding original vibration data, calculating the spectrum distribution through fast Fourier transform, comparing the energy ratio threshold of the standard frequency band, marking the abnormal frequency band, and generating the frequency response correlation path; S5: According to the starting node number of the frequency response association path, the device topology database is retrieved to obtain the corresponding component code, and the valve opening setting value and the pump speed feedback value are compared in real time. When the deviation exceeds the preset tolerance, the fault component number is output.

Images