Fault self-diagnosis and isolation method for sensors of weighing system of concrete mixing plant

Abstract

This invention discloses a self-diagnosis and isolation method for sensor faults in a concrete batching plant weighing system, belonging to the field of industrial automatic control and fault diagnosis technology. Addressing the problem of traditional systems failing to identify individual sensor faults due to parallel connection of multiple sensor signals and lacking online isolation, leading to production interruptions, this method connects an independent analog-to-digital converter to the output of each weighing sensor and installs a reference sensor. During the material's falling impact, the transient response waveforms of each sensor are captured synchronously, and peak value and attenuation rate characteristics are extracted. The current characteristics are compared with the sensor's own historical characteristics and those of other sensors; if the deviation exceeds a threshold, a fault is determined. Faulty sensors are isolated using electronic switches on the signal path. The total weight of the material is calculated using a radial basis function neural network based on the outputs of the remaining normal sensors. This method is used for self-diagnosis and isolation of faults in concrete batching plant weighing systems, allowing for continuous operation without shutdown even when a sensor fails.

Description

Technical Field

[0001] This invention belongs to the field of industrial automatic control and fault diagnosis technology, specifically relating to a method for self-diagnosis and isolation of sensor faults in a concrete mixing plant weighing system. Background Technology

[0002] Traditional concrete batching plant weighing systems typically connect the output signals of multiple weighing sensors in parallel to a single analog-to-digital converter (ADC). The controller can only obtain the sum of all sensor signals. When one sensor fails, although the overall output signal will be abnormal, the system cannot determine which sensor is faulty, nor can it distinguish between sensor failure and normal operating disturbances such as material impact. Because the faulty sensor cannot be located, operators often need to shut down the machine and check each sensor individually, leading to production interruptions. Furthermore, existing systems lack online isolation mechanisms after sensor failure; the faulty sensor remains connected in the signal path, and its abnormal output continues to affect the reliability of the weighing results. Simultaneously, the system cannot continue weighing using the remaining working sensors; production can only resume after the fault is resolved. These limitations make the continuous operation capability and maintainability of the weighing system insufficient to meet the actual production needs of concrete batching plants. Summary of the Invention

[0003] One object of the present invention is to address at least the aforementioned deficiencies and to provide at least the advantages that will be described later.

[0004] This invention provides a method for self-diagnosis and isolation of sensor faults in a concrete mixing plant weighing system. It can identify a single faulty sensor in the weighing system, isolate the faulty sensor from the signal path, and continue to calculate the material weight using the output signals of the remaining normal sensors.

[0005] This invention provides a method for self-diagnosis and isolation of sensor faults in a concrete mixing plant weighing system, comprising the following steps: S1: Connect an independent analog-to-digital converter to the output of each weighing sensor in the concrete mixing plant to convert the analog voltage signal output by each weighing sensor into a digital signal and input it into the controller respectively; S2: A reference sensor is installed in the weighing system near the point of impact of the material falling. The output of the reference sensor is connected to the controller via an independent analog-to-digital converter. S3: During the dynamic process of material falling from the silo to the metering hopper, the controller synchronously captures the transient response waveforms of each weighing sensor at the moment the material hits the metering hopper, and extracts the peak characteristics and decay rate characteristics of the transient response waveforms. S4: The controller compares the peak characteristics of the current transient response waveform of the same weighing sensor with the peak characteristics of the historical transient response waveform of the same weighing sensor. If the deviation exceeds the first preset deviation threshold, a longitudinal peak anomaly is recorded. The controller compares the decay rate characteristics of the current transient response waveform of the same weighing sensor with the decay rate characteristics of the historical transient response waveform of the same weighing sensor. If the deviation exceeds the third preset deviation threshold, a longitudinal decay anomaly is recorded. The controller compares the peak characteristics of the current transient response waveform of a weighing sensor with the average peak value of the current transient response waveforms of other weighing sensors and the peak value of the current transient response waveform of a reference sensor. If the deviation exceeds the second preset deviation threshold, a lateral anomaly is recorded. When longitudinal peak anomaly, longitudinal decay anomaly, and lateral anomaly all occur, the weighing sensor is determined to be faulty. S5: An electronic switch is connected in series in the signal transmission path between each weighing sensor and its corresponding analog-to-digital converter. When the controller determines that a weighing sensor has failed, it sends a disconnect command to the electronic switch corresponding to that weighing sensor to hardware isolate the faulty sensor from the weighing system. S6: The controller uses a pre-trained radial basis function neural network to calculate the total weight of the material based on the output signals of the remaining normal weighing sensors.

[0006] Preferably, in step S3, the controller synchronously captures the transient response waveforms of each weighing sensor at the instant the material impacts the weighing hopper, specifically including the following steps: S301: The controller uses the voltage change rate detected by the reference sensor as the synchronous trigger source. When the voltage change rate exceeds the preset slope trigger threshold, the controller simultaneously sends a trigger sampling command to the analog-to-digital converters corresponding to all weighing sensors. S302: The analog-to-digital converter corresponding to each weighing sensor responds to the trigger sampling command and synchronously acquires the voltage signal sequence of the weighing sensor during the impact period at the same sampling frequency; S303: The controller performs DC bias removal and filtering on the voltage signal sequence collected by each weighing sensor to obtain the original transient response waveform of each weighing sensor. S304: The controller performs amplitude normalization processing on the original transient response waveforms of each weighing sensor. Amplitude normalization processing is to divide the voltage amplitude of each sampling point in the original transient response waveform of each weighing sensor by the reference impulse response peak value recorded by the weighing sensor during the system calibration phase. S305: The controller extracts peak characteristics and decay rate characteristics from the transient response waveform after amplitude normalization.

[0007] Preferably, in step S303, the controller performs DC bias removal and filtering on the voltage signal sequence acquired by each weighing sensor to obtain the original transient response waveform of each weighing sensor, specifically including the following steps: S3031: When the material is not falling, the controller continuously records the voltage signal sequence of each weighing sensor and the reference sensor within the second monitoring time window, calculates the average amplitude of the voltage signal sequence of each weighing sensor and the reference sensor within the second monitoring time window, and uses the average amplitude as the dynamic DC bias value of the corresponding weighing sensor and the reference sensor in the current impact event. S3032: The controller subtracts the dynamic DC bias value corresponding to the weighing sensor from the voltage amplitude of each sampling point in the voltage signal sequence collected by each weighing sensor during the impact period to obtain the voltage signal sequence after DC bias removal. S3033: The controller uses the voltage signal sequence of the reference sensor after DC bias removal as the reference input signal of the adaptive filter, and uses the voltage signal sequence of each weighing sensor after DC bias removal as the desired response signal of the adaptive filter. S3034: The adaptive filter iteratively updates the filter weight coefficients based on the error signal between the reference input signal and the desired response signal until the mean square value of the error signal converges to a preset range. The updated filter weight coefficients are then applied to the filtered output signal generated by the reference input signal, which serves as the original transient response waveform of the weighing sensor.

[0008] Preferably, in step S305, the controller extracts peak characteristics and decay rate characteristics from the transient response waveform after amplitude normalization, specifically including the following steps: S3051: The controller performs Hilbert transform on the transient response waveform data after amplitude normalization to obtain the envelope curve of the transient response waveform; S3052: The controller searches for the maximum amplitude point in the envelope curve and uses the normalized voltage amplitude corresponding to the maximum amplitude point as the peak characteristic of the transient response waveform; S3053: The controller takes the sampling time corresponding to the maximum amplitude as the starting point, and extracts the attenuation segment data of the envelope curve along the time axis. It then uses the least squares method to fit the attenuation segment data of the envelope curve to an exponential function. The fitting model expression is y(t)=A▪e (-t / τ) , where y(t) is the amplitude of the envelope curve at time t, A is the initial amplitude parameter for fitting, and τ is the decay time constant parameter; S3054: The controller extracts the fitted decay time constant parameter τ as the decay rate characteristic of the transient response waveform.

[0009] Preferably, in step S4, the specific steps for the controller to perform feature comparison include: S401: The controller stores the operating condition parameter group corresponding to each impact event and the peak value and decay rate characteristics extracted by each weighing sensor in that impact event. The operating condition parameter group includes the material type identifier, hopper opening parameter and the initial material weight value of the metering hopper. S402: For the current impact event, the controller retrieves the K stored historical impact events based on the current operating condition parameter set, where K is an odd number, and the Euclidean distance between the operating condition parameter set of the K historical impact events and the current operating condition parameter set is minimized. S403: The controller uses the weighted average of the peak characteristics in the K historical impact events as the adaptive peak reference of the weighing sensor under the current impact event, and the weighted average of the attenuation rate characteristics in the K historical impact events as the adaptive attenuation rate reference of the weighing sensor under the current impact event, wherein the weighting coefficient is inversely proportional to the Euclidean distance. S404: The controller calculates the deviation between the current peak characteristic of the weighing sensor and the adaptive peak reference. If the deviation exceeds the first preset deviation threshold, a longitudinal peak abnormality is recorded. The controller also calculates the deviation between the current attenuation rate characteristic of the weighing sensor and the adaptive attenuation rate reference. If the deviation exceeds the third preset deviation threshold, a longitudinal attenuation abnormality is recorded. S405: The controller calculates the average peak value of the current peak characteristics of the other normal weighing sensors in the current impact event, excluding the sensor being evaluated, calculates the ratio of the average peak value to the current peak characteristic of the reference sensor, uses the ratio as the lateral threshold scaling factor, and multiplies it with the second preset deviation threshold to obtain the adaptive lateral deviation threshold. S406: The controller calculates the deviation between the current peak characteristic of the weighing sensor and the average current peak characteristic of other normal weighing sensors. If the deviation exceeds the adaptive lateral deviation threshold, a lateral anomaly is recorded. S407: The controller confirms that the weighing sensor has malfunctioned if it has already recorded a longitudinal peak abnormality based on the longitudinal peak deviation exceeding the first preset deviation threshold, a longitudinal attenuation abnormality based on the longitudinal attenuation deviation exceeding the third preset deviation threshold, and a lateral abnormality based on the lateral peak deviation exceeding the adaptive lateral deviation threshold. S408: If only one or two deviations exceed the corresponding threshold, the controller will maintain the working state of the weighing sensor and trigger an early warning signal, while increasing the sampling frequency and comparison accuracy of subsequent impact events.

[0010] Preferably, in step S6, the controller calculates the total weight of the material using a pre-trained radial basis function neural network, specifically including: S601: The controller pre-establishes radial basis function neural network sub-models corresponding to each combination of working conditions, including the combination of working conditions where all weighing sensors in the weighing system are normal and the combination of working conditions where all normal weighing sensors remain after at least one weighing sensor is isolated. The number of input nodes in each sub-model is equal to the number of normal weighing sensors in the corresponding combination of working conditions. S602: The controller uses the sample set consisting of the known material weight and the corresponding steady-state output voltage signal of the normal weighing sensor collected under each combined working condition to train the corresponding radial basis function neural network sub-model, and stores each trained sub-model as a model library with the combined working condition as the index. S603: During online weighing, after the controller completes hardware isolation of the faulty sensor, the controller identifies the combined working conditions composed of the remaining normal weighing sensors and retrieves the radial basis function neural network sub-model that matches the combined working conditions from the model library. S604: The controller inputs the steady-state voltage output signal of the remaining normal weighing sensor to the retrieved sub-model, and uses the output value of the sub-model as the total weight value of the material.

[0011] Preferably, in step S601, the controller pre-establishes radial basis function neural network sub-models corresponding to each normal sensor combination condition, specifically including: S6011: The controller calculates the geometrically invariant characteristic parameters of the remaining normal weighing sensor position distribution under each normal sensor combination working condition based on the installation position coordinates of each weighing sensor at the bottom of the weighing hopper. S6012: The controller determines different combinations of working conditions where the weighted Euclidean distance between geometrically invariant characteristic parameters is less than a preset equivalent distance threshold as mechanically equivalent combinations of working conditions, and divides all normal sensor combinations of working conditions into several mechanically equivalent groups. S6013: The controller trains a radial basis function neural network sub-model for each mechanical equivalence group. This sub-model is used as the sub-model corresponding to each combination of working conditions within the mechanical equivalence group. The sub-model uses the union sample set of known material weights collected under all combination working conditions within the mechanical equivalence group and the steady-state output voltage signal of the corresponding normal weighing sensor as training data, and establishes a mapping relationship from each combination of working conditions to the corresponding sub-model of the mechanical equivalence group.

[0012] Preferably, in step S6011, the calculation of the geometrically invariant characteristic parameters specifically includes: S60111: During the system calibration phase, the controller sequentially applies a unit calibration load to each weighing sensor mounting point, records the output voltage response of each weighing sensor, and constructs an N×N dimensional unit load influence coefficient matrix, where N is the total number of weighing sensors, and the matrix elements represent the output response value of the i-th sensor when the load is applied at the j-th mounting point. S60112: For any normal sensor combination working condition, the controller extracts an M×M dimensional submatrix from the unit load influence coefficient matrix, which is composed of the rows and columns corresponding to the remaining normal weighing sensors in the combined working condition, where M is the number of remaining normal weighing sensors in the combined working condition. S60113: The controller performs singular value decomposition on the M×M dimensional submatrix, sorts the resulting M singular values ​​in descending order and divides them by M to form the normalized singular value vector of the combined working condition, and uses the normalized singular value vector as the geometrically invariant characteristic parameter.

[0013] The present invention has at least the following beneficial effects: First, this invention connects an independent analog-to-digital converter to the output of each weighing sensor, allowing the controller to independently acquire the output data of each sensor, thus providing a basis for fault location. During the material impact process, each sensor outputs a transient response waveform. The peak value and decay rate characteristics of this waveform are related to the sensor's own state. The controller compares the current waveform characteristics with its own historical characteristics and the current characteristics of other sensors to determine whether a sensor has malfunctioned. When a sensor is determined to be faulty, the controller disconnects it via an electronic switch connected in series in the signal path, achieving hardware isolation of the faulty sensor and preventing its abnormal output from affecting the weighing results. Based on this, the controller uses a radial basis function neural network to calculate the total weight of the material in real time based on the output signals of the remaining normal sensors. In this way, even if a sensor malfunctions, the system can still isolate the fault online and continue operating without downtime for maintenance, thereby improving the continuous operation capability and maintainability of the weighing system.

[0014] Secondly, this invention connects an independent analog-to-digital converter to each weighing sensor signal path and uses the voltage change rate detected by the reference sensor as a synchronous trigger source. All weighing sensors can simultaneously initiate sampling at the moment of material impact, ensuring strict temporal synchronization of the transient response waveforms captured by each sensor. Subsequently, the voltage signal sequences acquired by each sensor undergo DC bias removal and filtering, and the processed waveforms are normalized by dividing by the peak value of each sensor's own reference impact response recorded during the calibration phase. Thus, even if different sensors have different original response amplitudes due to variations in installation location, the normalized waveform characteristics are comparable, providing a unified data foundation for subsequent fault diagnosis based on waveform characteristics.

[0015] Third, during signal processing, the controller continuously records the average voltage signal of each sensor within a fixed time window as a dynamic DC bias when the material is not falling. This bias is subtracted during the impact period to eliminate the interference of temperature drift and time-varying factors on the baseline level. Simultaneously, the DC-biased signal from the reference sensor is used as the reference input to the adaptive filter, and the DC-biased signal from each weighing sensor is used as the desired response. By iteratively updating the filter weight coefficients, the filter output cancels out the common-mode vibration components related to the reference signal. The original transient response waveform obtained after adaptive filtering retains the true characteristics of the impact excitation, while background vibration interference superimposed on the signal is effectively suppressed.

[0016] Fourth, from the transient response waveform after amplitude normalization, the controller first extracts the envelope curve of the waveform through Hilbert transform, and searches for the point with the maximum amplitude on the envelope curve as the peak feature. Then, taking the time corresponding to the peak point as the starting point, the decay segment of the envelope curve is extracted, and the least squares method is used to fit an exponential function to this decay segment. The decay time constant obtained by fitting is used as the decay rate feature. Compared with directly using the maximum amplitude of discrete sampling points or a fixed proportion decay interval, the envelope extraction and exponential fitting method can eliminate the influence of sampling noise and local waveform fluctuations on the feature values. Moreover, the decay time constant reflects the damping characteristics of the sensor's elastic body and mechanical structure, and has an adaptive normalization effect for different impact intensities, making the extracted features stable and comparable under different working conditions.

[0017] Fifth, the controller stores the operating parameters corresponding to each impact event, such as material type, silo opening, and initial weight of the metering hopper, along with the extracted peak value and decay rate features. For the current impact event, the controller retrieves the K events with the smallest Euclidean distance from the historical operating parameters, uses the weighted average of their peak value and decay rate features as an adaptive benchmark, and calculates the deviation between the current feature and the benchmark. If the deviation exceeds the first or third preset threshold, a longitudinal anomaly is recorded. Simultaneously, the second preset deviation threshold is scaled using the ratio of the current average peak value of the remaining normal sensors to the current peak value of the reference sensor to obtain an adaptive lateral deviation threshold. The deviation between the current sensor peak value and the average peak value of the remaining sensors is compared with this threshold; if it exceeds, a lateral anomaly is recorded. Only when all three parameters—longitudinal peak deviation, longitudinal decay deviation, and lateral deviation—exceed the corresponding threshold is the sensor confirmed to be faulty. If only one or two parameters exceed the threshold, the sensor remains operational and an early warning is issued, while simultaneously increasing the sampling frequency and comparison accuracy for subsequent impact events. This multi-criteria mechanism, based on adaptive operating conditions and dynamic threshold scaling, can avoid false alarms and false isolations caused by fluctuations in normal operating conditions or accidental deviations in the material drop point.

[0018] Sixth, after isolating the faulty sensor, the controller needs to recalculate the total weight of the material using the outputs of the remaining normal sensors. To this end, radial basis function neural network sub-models with the number of input nodes equal to the number of sensors in each combination are pre-built for all possible combinations of sensors in the weighing system, including combinations where all sensors are functioning normally and combinations where at least one sensor is isolated. Each sub-model is trained using a sample set consisting of the known material weights collected for each combination and the corresponding steady-state output voltage signals of the sensors. During online operation, the controller retrieves the matching sub-model from the model library based on the current operating conditions of the remaining normal sensors, inputs the steady-state voltage outputs of each normal sensor into the sub-model, and uses its output value as the total weight of the material. In this way, regardless of which sensor is isolated, the neural network input dimension always strictly corresponds to the number of currently effective sensors, and the model parameters are perfectly adapted to the force transmission characteristics between the remaining sensors, thus maintaining weighing accuracy comparable to the full sensor state under different fault isolation conditions.

[0019] Seventh, when the number of sensors is large, training independent sub-models for each possible combination of normal sensors leads to combinatorial explosion, causing a sharp increase in training sample size, training time, and storage space. To address this, the controller calculates the geometrically invariant characteristic parameters of the remaining normal sensor positions under each combination of working conditions based on the installation coordinates of each weighing sensor at the bottom of the weighing hopper. Combinations with a weighted Euclidean distance less than a preset threshold are classified as mechanically equivalent combinations and grouped into the same mechanically equivalent group. Each mechanically equivalent group shares a radial basis function neural network sub-model. The training data is the union of sample sets collected under all combinations of working conditions within the group, and a mapping relationship is established from each combination of working conditions to its corresponding shared sub-model. Through this merging strategy, the number of sub-models that need to be trained and stored is significantly reduced without substantially sacrificing weighing accuracy, improving the engineering practicality of the method.

[0020] Eighth, to accurately reflect the local stiffness differences at the bottom of the weighing hopper and the mechanical coupling effects between the sensors, the geometrically invariant characteristic parameters are not directly calculated based on the installation position coordinates. Instead, during the system calibration phase, a unit calibration load is applied at each point, and the output voltage response of each sensor is recorded to construct a unit load influence coefficient matrix. For any normal sensor combination, a submatrix consisting of the rows and columns corresponding to the remaining normal sensors is extracted from this matrix. Singular value decomposition is performed on this submatrix, and the resulting singular values ​​are sorted from largest to smallest and divided by the number of sensors to form a normalized singular value vector, which is used as the geometrically invariant characteristic parameter. This vector comprehensively characterizes the force transmission gain characteristics, anisotropy, and local stiffness differences under this sensor layout from the perspective of system energy. This ensures that the division of the mechanically equivalent group takes into account both geometric distribution and mechanical coupling factors, thereby compressing the number of sub-models while maintaining the weighing accuracy of the shared sub-models under various merged conditions.

[0021] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to embodiments, so that those skilled in the art can implement it based on the description.

[0023] Traditional concrete batching plant weighing systems connect the output signals of multiple weighing sensors in parallel to a single analog-to-digital converter channel. The controller can only obtain the sum of all sensor signals and cannot acquire the independent output of a single sensor. When a sensor malfunctions, the system can only detect the abnormality in the overall signal but cannot pinpoint the specific faulty sensor or distinguish between sensor failure and normal operating disturbance.

[0024] Because the faulty sensor could not be located, operators had to shut down the machine and check each sensor one by one, causing production interruptions. The faulty sensor remained connected in the signal path, and its abnormal output continued to affect the reliability of the weighing results. The system could not continue weighing using the remaining working sensors and could only resume production after the fault was resolved. These limitations made it difficult for the weighing system to meet actual production needs in terms of continuous operation and maintainability.

[0025] To address the aforementioned problems, this embodiment provides a self-diagnosis and isolation method for sensor faults in a concrete batching plant weighing system, implemented as follows: First, an independent analog-to-digital converter (ADC) is connected to the output terminal of each weighing sensor in a concrete batching plant. The analog voltage signal output by each sensor is converted into a digital signal and then input to a programmable logic controller (PLC). Simultaneously, a reference sensor is installed near the impact point of the falling material in the weighing system. The output of this reference sensor is also connected to the same controller via an independent ADC. An electronic switch is connected in series in the signal transmission path between each weighing sensor and its corresponding ADC. The on / off state of this switch is controlled by the controller. During system operation, when material falls from the hopper into the weighing hopper, the controller uses the voltage change rate detected by the reference sensor as a synchronous trigger source. Once the voltage change rate exceeds a preset slope threshold, the controller immediately sends a trigger sampling command to all ADCs corresponding to the weighing sensors simultaneously. Each ADC synchronously acquires the voltage signal sequence during the impact period at the same sampling frequency. The controller first continuously records the average voltage of each weighing sensor within a fixed time window while the material is not falling, using this as a dynamic DC bias. This bias is then subtracted from the voltage sequence collected during the impact period. The debiased signal from the reference sensor is used as the reference input for an adaptive filter, and the debiased signal from each weighing sensor is used as the desired response. The filter weight coefficients are iteratively updated, and the filtered output is used as the original transient response waveform. The controller performs amplitude normalization on the original transient response waveforms of each sensor, i.e., dividing it by the reference impact response peak value recorded during the calibration phase. Then, peak characteristics and decay rate characteristics are extracted from the normalized waveforms. The controller compares the peak characteristics of the current impact event for the same sensor with the peak characteristics of historical impact events for that sensor. If the deviation exceeds a first preset deviation threshold, a longitudinal peak anomaly is recorded. The current decay rate characteristic is compared with historical decay rate characteristics; if the deviation exceeds a third preset deviation threshold, a longitudinal decay anomaly is recorded. Simultaneously, the current sensor peak characteristics are compared with the average peak values ​​of other normal sensors and the reference sensor peak value; if the deviation exceeds a second preset deviation threshold, a lateral anomaly is recorded. When longitudinal peak abnormality, longitudinal attenuation abnormality, and lateral abnormality all occur, the sensor is determined to be faulty. When all three abnormalities occur simultaneously, the controller determines the sensor is faulty and immediately sends a disconnect command to the corresponding electronic switch, hardware isolating the faulty sensor from the weighing system. Finally, the controller invokes a pre-trained radial basis function neural network, which takes the steady-state output voltage signal of the remaining normal sensors as input and outputs the total weight value of the material.

[0026] After adopting the method of this embodiment, actual tests conducted at a concrete mixing plant showed that, in the test simulating a single weighing sensor failure, the system's accuracy in identifying the faulty sensor reached 99.2%, and the average time from the occurrence of the fault to the completion of hardware isolation was less than 45 milliseconds. After isolating the faulty sensor, the maximum relative error between the total material weight calculated using the remaining three normal sensors and the radial basis function neural network and the actual weighing value was ±0.47%, while the traditional parallel system could only shut down for maintenance under the same fault conditions, with an average downtime of 32 minutes. The above data indicate that this method can effectively identify and isolate faulty sensors, achieving continuous operation without shutdown, and significantly improving the reliability and production efficiency of the weighing system.

[0027] Building upon the methods described above, simultaneously capturing and comparing the transient response waveforms of each weighing sensor is crucial for accurate fault diagnosis. Traditionally, each sensor is triggered independently or with a fixed time delay. Due to random fluctuations in the impact time of the falling material, the waveforms captured by different sensors are difficult to align in time, resulting in inconsistent starting points and phases of the response waveforms under the same impact event, making effective lateral comparison impossible. Furthermore, because different sensors are installed at varying distances from the connection point to the weighing hopper, their response amplitudes differ significantly even under the same impact excitation. Directly comparing the original amplitudes can mistakenly interpret differences in installation location as sensor faults.

[0028] To address the aforementioned issues, this embodiment further provides the following specific solution based on the above implementation method. The controller uses the voltage change rate detected by the reference sensor as a unified synchronous trigger source. Specifically, the controller continuously monitors the output voltage of the reference sensor and calculates its first derivative as the voltage change rate in real time. When material falls from the hopper and impacts the weighing hopper, the reference sensor output signal exhibits a steep rising edge, and its voltage change rate increases rapidly. The controller pre-sets a slope trigger threshold, which is slightly higher than the maximum change rate caused by normal vibration noise. When the detected voltage change rate exceeds this threshold, the controller immediately sends a trigger sampling command to the analog-to-digital converters (ADCs) corresponding to all weighing sensors simultaneously. After receiving the command, each ADC synchronously acquires the voltage signal sequence of its connected sensor during the impact period at the same preset sampling frequency, such as 10 kHz. The acquisition duration covers the entire process from the start of the impact to the waveform decaying to a stable state, typically set to 200 milliseconds. Because all ADCs share the same trigger source and start simultaneously, the transient response waveforms captured by each sensor are strictly aligned on the time axis, and the deviation of the impact start point is less than one sampling period.

[0029] After obtaining the synchronously acquired raw voltage sequence, the controller performs DC bias removal and filtering on each signal to obtain the raw transient response waveforms of each weighing sensor. Then, the controller performs amplitude normalization on the raw transient response waveforms of each sensor. During system calibration, when the weighing hopper is empty and there is no material impact, the controller applies a standard impact force at the material drop point through an external excitation device, records the response waveform of each weighing sensor under this standard impact, and uses its peak voltage as the reference impact response peak value for that sensor. During normal production, the controller divides the voltage amplitude at each sampling point in the raw transient response waveform of each weighing sensor by the reference impact response peak value pre-stored by that sensor to obtain the normalized waveform. In this way, even if a sensor's raw response amplitude is only half that of a nearby sensor due to its distant installation location, the peak value of each sensor under the same impact intensity after normalization will be close to 1, eliminating the fixed deviation caused by differences in installation location. The peak characteristics and attenuation rate characteristics extracted from the normalized waveforms then provide comparability between sensors, providing a unified data basis for subsequent fault diagnosis.

[0030] After adopting the above scheme, comparative tests were conducted at the same concrete mixing plant. Without synchronous triggering, the maximum time difference between the starting points of the waveforms captured by different sensors reached 12 milliseconds, resulting in a peak deviation fluctuation range of 15% for each sensor under the same impact. After synchronous triggering using the reference sensor voltage change rate, the starting point time difference was less than 0.1 milliseconds, and the peak deviation fluctuation range was reduced to within 2%. Regarding amplitude normalization, the same impact force was applied to four weighing sensors installed at different locations, with original peak voltages of 1.2 V, 0.9 V, 0.6 V, and 0.4 V, respectively. The maximum difference in direct comparison was 200%. After normalization, the normalized peak values ​​of the four sensors were all between 0.98 and 1.02, with a deviation of less than 4%. This indicates that the method effectively achieves synchronous capture and amplitude comparability of multi-sensor transient waveforms, laying a reliable foundation for subsequent accurate fault diagnosis based on waveform characteristics.

[0031] Based on the above implementation method, obtaining the original transient response waveform before synchronous triggering and amplitude normalization still faces two prominent problems. First, the output voltage of the weighing sensor has a slowly changing DC bias over time. This is caused by zero-point drift due to changes in ambient temperature, amplifier temperature drift, and static load changes caused by the accumulation of residual material in the weighing hopper. These DC biases are not the same in each impact event. If a fixed bias value is directly subtracted, the baseline cannot be zeroed, resulting in distortion of the extracted peak value and decay rate. Second, during the material falling impact, other equipment in the mixing plant, such as vibrating motors and belt conveyors, will generate broadband mechanical vibrations. These vibrations are coupled to all weighing sensors through the weighing hopper structure, forming common-mode interference. The frequency band of this interference severely overlaps with the frequency band of the impact response signal. Conventional filters cannot separate it from the useful signal, and forced filtering will damage the steep edge characteristics of the impact response.

[0032] To address the aforementioned issues, this embodiment further provides the following specific solution based on the above implementation method. First, dynamic DC bias is addressed. Before each impact event occurs, i.e., during the static period before the material falls from the hopper, the controller continuously records the output voltage of each weighing sensor. Specifically, the controller sets a second monitoring time window with a length of 100 milliseconds, located after the previous impact response has completely decayed and before the next material dropping command is issued. Within this window, the controller collects the voltage value of each weighing sensor at a rate of 1,000 times per second and calculates the arithmetic mean of these sampling points. This average value is used as the dynamic DC bias value of the sensor in the current impact event. When the material falls and impact occurs, the controller subtracts this dynamic DC bias value from each voltage sampling point collected during the impact period to obtain a voltage signal sequence after DC bias removal. Since the bias is recalculated before each impact, the effects of temperature drift and time-varying loads are eliminated successively.

[0033] After DC bias removal, the controller adaptively cancels common-mode vibration interference. The controller uses the de-DC biased voltage signal sequence of the reference sensor as the reference input signal for the adaptive filter, and the de-DC biased voltage signal sequence of each weighing sensor as the desired response signal of the adaptive filter. The adaptive filter employs the least mean square algorithm, with a transverse filter structure and an order of 32. The filtering process is as follows: the filter calculates the output signal based on the current reference input signal and the filter weight coefficients, and subtracts this output signal from the desired response signal to obtain the error signal; based on the error signal and the reference input signal, the filter weight coefficients are updated according to the iterative formula of the least mean square algorithm, with a step size factor of 0.01; this process is repeated until the mean square value of the error signal changes less than a preset convergence threshold (e.g., 1e-6) over 50 consecutive iterations. At this point, the filter weight coefficients have converged to their optimal value, and the signal output by the filter is the best estimate of the common-mode interference related to the reference input. This interference estimate is subtracted from the desired response signal to obtain the original transient response waveform. Since the reference sensor is installed near the impact point of the material drop, its response to common-mode vibration interference is highly correlated with that of other weighing sensors, while its response to local impact varies depending on the installation location. Therefore, the adaptive filter can effectively separate and cancel the common-mode vibration component, while retaining the unique impact response waveform of each weighing sensor.

[0034] After adopting this scheme, a field test was conducted at a concrete mixing plant. Without dynamic DC bias removal, in 10 consecutive impact events, the DC bias drift of the same sensor reached 15% of the initial bias value, resulting in a peak feature fluctuation of ±8%. After using dynamic bias calculation, the baseline offset was controlled within 0.5% of the initial bias value, and the peak feature fluctuation was reduced to ±0.7%. Regarding common-mode vibration suppression, when the mixing plant's vibrating motor was running, the impact response waveform was completely submerged in vibration noise in the original signal without adaptive filtering, with a signal-to-noise ratio of only 3 dB. After adaptive filtering cancellation, the impact response waveform was clearly distinguishable, the signal-to-noise ratio improved to 18 dB, and the steepness of the waveform's starting edge remained intact, without edge broadening caused by filtering. The above data show that this method can effectively eliminate dynamic DC bias and common-mode vibration interference, providing a high-quality signal foundation for subsequent extraction of stable and reliable transient response waveform features.

[0035] Based on the above implementation method, when extracting peak and decay rate features from the amplitude-normalized transient response waveform, the traditional approach is to directly find the maximum value among the discrete sampling points as the peak value, and use the time interval corresponding to the waveform decreasing from the peak value to a certain fixed ratio, such as 50%, as the decay rate. This method has obvious drawbacks: First, discrete sampling points are affected by high-frequency noise, and the maximum value may deviate from the true peak position, resulting in large fluctuations in peak features; second, using the time interval corresponding to a fixed amplitude ratio as the decay rate is not stable under different impact intensities, because the initial amplitude of the waveform changes when the impact intensity changes, but the essential time constant of the decay process does not change. The fixed ratio method will incorrectly give different decay rate values, making the decay features between different impact events incomparable. In addition, small local oscillations in the waveform can also interfere with peak location and decay segment truncation.

[0036] To address the aforementioned problems, this embodiment further provides the following specific solution based on the above implementation method. The controller first performs a Hilbert transform on the transient response waveform data after amplitude normalization. Specifically, let the normalized waveform sequence be x[n], n=0,1,…,L-1, where L is the sequence length. The controller calculates the Hilbert transform of x[n] to obtain the imaginary part of its analytic signal, and then obtains the envelope curve e[n]=sqrt(x[n]). 2 +(Hilbert(x[n])) 2 The Hilbert transform can be implemented in the frequency domain via the Fourier transform: x[n] undergoes a fast Fourier transform, with positive frequency components multiplied by -j and negative frequency components multiplied by j, followed by an inverse transform to obtain the imaginary part. The resulting envelope curve e[n] eliminates high-frequency carrier oscillations and exhibits a smooth, single-peak shape, facilitating subsequent feature extraction.

[0037] After obtaining the envelope curve, the controller searches for the point with the maximum amplitude in e[n] and records the index n corresponding to that point. peak and amplitude e max , will e max As the peak characteristic of the transient response waveform, since the envelope curve has been smoothed by Hilbert transform, the peak location is not affected by local sampling glitches in the original waveform.

[0038] Then, the controller extracts the decay rate features. This is done using the sampling time n corresponding to the peak point. peakStarting from the time axis, the attenuation segment data of the envelope curve is truncated, covering the time required for the value to decrease from the peak to 5% of the peak value. Typically, 100 sampling points are used. The controller uses the least squares method to fit an exponential function to the truncated attenuation segment data. The fitting model is y = A * exp(-t / τ), where y is the envelope amplitude, t is time, A is the initial amplitude parameter, and τ is the attenuation time constant parameter. To facilitate linear least squares solution, the natural logarithm of both sides of the model is taken, resulting in ln(y) = ln(A)-t / τ. Let Y = ln(y), c = ln(A), k = -1 / τ, then the model becomes Y = c + k·t. The controller uses the truncated attenuation segment data points (t… i ,y i ), calculate ln(y) i Then, the slope k and intercept c are solved using the standard least squares method, resulting in τ = -1 / k. During fitting, points with amplitudes below 5% of the envelope peak value are removed to avoid noise-dominated tail data affecting fitting accuracy. Finally, the fitting-derived decay time constant τ is used as the decay rate characteristic of the transient response waveform.

[0039] After adopting this scheme, comparative tests were conducted at the same concrete mixing plant. Gaussian white noise with a signal-to-noise ratio of 15 dB was superimposed on the waveform, and peak features were extracted using both the direct maximum method and the envelope peak method of this invention. In 100 repeated tests, the relative standard deviation of the peak extracted values ​​using the direct maximum method was 6.8%, while the relative standard deviation of the envelope peak method was only 1.2%. For the decay rate feature, three impacts of different intensities were applied to the same sensor, and the actual decay time constant τ was 25 milliseconds. The results obtained by directly using the fixed 50% decay interval method were 22 milliseconds, 26 milliseconds, and 30 milliseconds, with a relative deviation of 20%; while the τ values ​​extracted using exponential fitting were 24.8 milliseconds, 25.1 milliseconds, and 25.3 milliseconds, with a maximum relative deviation of only 1.2%. The above data show that this method can effectively suppress the influence of noise and impact intensity variations on feature extraction, obtaining stable and physically meaningful peak and decay rate features, providing reliable feature input for subsequent accurate fault diagnosis.

[0040] Based on the above implementation methods, after obtaining stable and comparable peak and decay rate characteristics, setting the comparison benchmark and deviation threshold becomes crucial for fault diagnosis. Traditional methods typically use a fixed benchmark and fixed threshold, i.e., pre-store a set of peak and decay rates under normal conditions as a benchmark and set a fixed allowable deviation range. However, the operating conditions of concrete mixing plants change frequently, with materials ranging from crushed stone to sand, hopper openings adjusted according to flow demand, and initial weights of materials already in the metering hopper varying. These changes in operating conditions cause significant drift in the peak and decay rates of the transient response waveform within the normal range. If a fixed benchmark is used, normal fluctuations caused by changes in operating conditions can easily be misjudged as sensor faults; if the threshold is amplified to encompass changes in operating conditions, genuine sensor faults may be missed. Furthermore, accidental offsets at the material drop point or flow fluctuations from a single impact can also cause normal dispersion in the peak distribution among different sensors, making it difficult to balance sensitivity and robustness with a fixed lateral threshold.

[0041] To address the aforementioned issues, this embodiment further provides the following specific solution based on the above implementation method. The controller first stores the operating condition parameter group corresponding to each impact event. The operating condition parameter group includes three dimensions: material type identifier (e.g., crushed stone coded as 1, sand coded as 2, cement coded as 3); hopper opening parameter, i.e., the percentage of the discharge valve open, ranging from 0 to 100%; and the initial material weight value in the metering hopper, i.e., the weight of the material already in the hopper before this discharge. Simultaneously, the controller stores the peak value characteristics and attenuation rate characteristics extracted by each weighing sensor during this impact event.

[0042] For a current impact event, the controller acquires the current set of operating parameters and then retrieves the K most similar historical impact events from the historical storage data. K is an odd number, such as 9. Similarity is measured by the Euclidean distance between the operating parameter sets; that is, the Euclidean distance between the current operating vector and each historical operating vector is calculated, and the K events with the smallest distances are selected. The controller uses the weighted average of the peak characteristics of these K events as the adaptive peak reference for the weighing sensor under the current impact event. The weighting coefficient is inversely proportional to the Euclidean distance, meaning that historical events with closer distances have greater weights. Similarly, the weighted average of the decay rate characteristics of these K events is used as the adaptive decay rate reference. In this way, the reference value automatically drifts with changes in operating conditions, eliminating the influence of normal operating condition fluctuations on deviation calculations.

[0043] The controller calculates the deviation between the current peak characteristic and the adaptive peak reference. If the deviation exceeds a first preset deviation threshold, such as 15% of the reference value, a longitudinal peak anomaly is recorded. At the same time, it calculates the deviation between the current decay rate characteristic and the adaptive decay rate reference. If the deviation exceeds a third preset deviation threshold, such as 12% of the reference value, a longitudinal decay anomaly is recorded.

[0044] For lateral comparison, the controller calculates the average current peak characteristic of all normal weighing sensors (excluding the sensor being evaluated) in the current impact event, and then calculates the ratio of this average to the current peak characteristic of the reference sensor. This ratio is used as a lateral threshold scaling factor and multiplied by a second preset deviation threshold to obtain an adaptive lateral deviation threshold. For example, when the ratio of the average peak value of the other normal sensors to the peak value of the reference sensor is 0.9, it indicates that the overall impact is weak, and the lateral deviation threshold is reduced accordingly; if the ratio is 1.1, it indicates that the impact is strong, and the threshold is increased accordingly. The controller calculates the deviation between the current peak characteristic of the sensor being evaluated and the average peak characteristic of the other normal sensors. If this deviation exceeds the aforementioned adaptive lateral deviation threshold, a lateral anomaly is recorded.

[0045] The controller employs a triple-criteria fault confirmation mechanism: a sensor is only confirmed to be faulty and hardware isolation is implemented when all three criteria—longitudinal peak deviation, longitudinal attenuation deviation, and lateral deviation—exceed their respective thresholds. If only one or two deviations exceed the threshold, the sensor remains operational without isolation, but a warning signal is triggered, and the sampling frequency and comparison accuracy for subsequent impact events are increased—from the normal sampling frequency of 10 kHz to 20 kHz. Deviation comparison is changed from single-event judgment to a comprehensive judgment based on three consecutive events to further observe the evolution trend of characteristics.

[0046] After adopting this solution, a week-long comparative test was conducted at a concrete mixing plant. Under frequent changes in operating conditions, the traditional method using a fixed benchmark and fixed threshold resulted in 17 false alarms and 2 missed alarms. However, using the adaptive benchmark and dynamic threshold method of this invention, the number of false alarms decreased to 1 and missed alarms were 0. Specifically, when the material changed from crushed stone to sand, the sensor peak characteristic normally decreased by 22%. The fixed benchmark method immediately triggered a false alarm, while the adaptive benchmark, by searching historical events of the nearest neighbor operating conditions, showed a synchronous decrease of 21.5% in the benchmark value, with a deviation of only 0.5%, and did not trigger an alarm. When a fault simulating a 30% decrease in sensor sensitivity was artificially created, the fixed threshold method failed to detect it due to an overly wide threshold setting. However, the proposed method, with its triple criterion, successfully detected the fault because the longitudinal peak deviation exceeded the 15% threshold and the lateral deviation also exceeded the threshold after scaling the ratio of the average values ​​of other sensors to the reference sensor. The above data demonstrates that this method can effectively adapt to changes in operating conditions and impact dispersion, significantly reducing the false alarm rate while maintaining high sensitivity, achieving reliable and accurate fault determination.

[0047] Based on the above implementation method, after diagnosing and isolating faulty sensors, the system needs to continue calculating the total weight of the material using the output signals of the remaining normal sensors. The traditional approach is to pre-train a fixed neural network model, where the number of input nodes equals the total number of sensors in the weighing system. When a sensor is isolated, the model still requires input signals from all sensors; for isolated sensors, only zero or the original value can be input, leading to a mismatch between the model's input dimensions and the actual number of effective sensors. Simultaneously, sensor isolation alters the force transmission path in the weighing hopper, changing the force coupling relationship between the remaining sensors. The fixed model's weights cannot adapt to this change, causing the weight calculation error to increase significantly with the number of faulty sensors. For example, when one of the four sensors is isolated, the weighing error using the traditional method may worsen from ±0.3% to over ±2.5%.

[0048] To address the aforementioned issues, this embodiment further provides the following specific solution based on the above implementation method. During the offline calibration phase of the system, the controller pre-establishes radial basis function neural network sub-models corresponding to each of the following possible combinations: one where all weighing sensors in the weighing system are functioning normally, and another where at least one weighing sensor is isolated, leaving only the remaining normal weighing sensors. The number of input nodes in each sub-model is strictly equal to the number of normal weighing sensors in the corresponding combination. For a system with four weighing sensors, all sensors functioning normally is one combination; isolating any one sensor results in four combinations; isolating any two sensors results in six combinations; and isolating any three sensors results in four combinations, for a total of fifteen combinations. For each combination, the controller applies a known weight of material to the weighing hopper and collects the steady-state output voltage signals of each normal sensor under that combination and the corresponding material weight, forming a training sample set. In each sample set, the input vector dimension is equal to the number of sensors in that combination, and the output is the total weight of the material. The controller uses a radial basis function neural network, with the number of hidden layer nodes set to twice the number of input nodes, a Gaussian radial basis function, and a linearly weighted output layer. Each sub-model is trained using a sample set. After training, each sub-model is stored as a model library with the combined working conditions as the index.

[0049] During online weighing, after the controller completes hardware isolation of the faulty sensor, it identifies the combined operating condition consisting of the remaining normal weighing sensors. For example, sensor 2 is isolated, while sensors 1, 3, and 4 are operating normally. Based on the identifier of this combined operating condition, the controller retrieves a matching radial basis function neural network sub-model from the model library. The controller reads the steady-state voltage output signals of the remaining normal sensors and inputs them into the retrieved sub-model in a preset order. The sub-model directly outputs the total weight of the material. Because the input dimension of the sub-model is completely consistent with the number of currently effective sensors, and the training data comes from the actual force transmission relationship under the same combined operating condition, the model can accurately reflect the coupling characteristics between the remaining sensors.

[0050] After adopting this scheme, tests were conducted on the weighing system of a four-sensor concrete mixing plant. Under normal sensor conditions, weighing 1000 kg of material using the corresponding sub-model resulted in a maximum relative error of ±0.28%. When simulated sensor 2 malfunctioned and was isolated, the traditional fixed-model method, which sets the input of sensor 2 to zero, resulted in a maximum relative error of ±2.7% when weighing the same material. However, this invention retrieves the sub-models corresponding to the remaining sensors 1, 3, and 4, which have an input dimension of 3. Weighing the same 1000 kg of material, the maximum relative error was ±0.32%, essentially the same as under normal conditions. Further testing with two isolated sensors showed that the error of the traditional method increased to ±5.1%, while the error of this method, which retrieves the sub-models corresponding to the two sensors, was only ±0.45%. These data demonstrate that this method can maintain weighing accuracy at a level comparable to the fully sensor-controlled state by switching the matching sub-model under any sensor malfunction and isolation condition, effectively solving the accuracy degradation problem caused by input mismatch and force coupling changes in the fixed-model approach.

[0051] Based on the above implementation method, an independent radial basis function neural network sub-model is trained for each possible normal sensor combination. When the number of sensors in the weighing system is large, a combinatorial explosion problem arises. For example, when the weighing system is equipped with 6 weighing sensors, the total number of possible normal sensor combinations reaches 2^6 - 1 = 63. If samples are collected, 63 sub-models are trained and stored one by one, a large amount of calibration time and storage space are required, which limits the practicality of engineering. At the same time, there may be cases where different combinations of working conditions have similar mechanical transmission characteristics, and training independent models separately will result in repetitive work.

[0052] To address the aforementioned issues, this embodiment further provides the following specific solution based on the above implementation method. The controller first calculates the geometrically invariant characteristic parameters of the remaining normal weighing sensor position distribution under each normal sensor combination condition based on the installation position coordinates of each weighing sensor at the bottom of the weighing hopper. These geometrically invariant characteristic parameters are used to quantify the similarity of the force transmission structure under different combination conditions. Specifically, during the system calibration phase, the controller acquires the installation position coordinates of each weighing sensor. For any normal sensor combination condition, it extracts the position coordinates of the remaining normal sensors and calculates invariant features such as the geometric center and eigenvalues ​​of the covariance matrix of these coordinate points, constructing a multidimensional vector as the geometrically invariant characteristic parameter.

[0053] Then, the controller classifies different combinations of working conditions where the weighted Euclidean distance between geometrically invariant feature parameters is less than a preset equivalence distance threshold as mechanically equivalent combinations, and divides all normal sensor combinations into several mechanically equivalent groups. For example, for a six-sensor system, if the weighted Euclidean distance between the position distribution feature vectors of the remaining sensors in combinations of isolated sensor 1 and isolated sensor 2 is less than the threshold, then these two combinations are considered mechanically equivalent and are assigned to the same mechanically equivalent group. The controller trains a shared radial basis function neural network sub-model for each mechanically equivalent group. The number of input nodes of this shared sub-model is equal to the number of normal weighing sensors in any combination of working conditions within the mechanically equivalent group. During training, the sample set consisting of the known material weights collected under all combinations of working conditions within the mechanically equivalent group and the steady-state output voltage signals of the corresponding normal weighing sensors is merged as training data. At the same time, the controller establishes a mapping relationship from each specific combination of working conditions to the corresponding shared sub-model of its mechanically equivalent group.

[0054] During online weighing, after the controller completes hardware isolation of the faulty sensor, it first identifies the combined working conditions composed of the remaining normal weighing sensors. Then, based on the mapping relationship, it finds the mechanical equivalent group to which the combined working condition belongs, retrieves the shared radial basis function neural network sub-model corresponding to the group, inputs the steady-state voltage output signal of the current normal sensor into the shared sub-model, and outputs the total weight value of the material.

[0055] After adopting this scheme, tests were conducted on a six-sensor concrete mixing plant weighing system. Traditional methods require training 63 sub-models, each requiring approximately 200 samples, totaling 12,600 samples. Training takes about 4 hours, and the model library occupies about 15 megabytes of storage space. However, using the mechanically equivalent combination merging method of this invention, the 63 combinations are divided into 9 mechanically equivalent groups, requiring only 9 shared sub-models to be trained. The total sample size is 1,800 samples, training takes about 35 minutes, and the model library occupies about 2.1 megabytes of storage space. Regarding weighing accuracy, for any sensor isolation combination, using a shared sub-model to weigh 1000 kg of material results in a maximum relative error of ±0.41%, while using an independent sub-model results in an error of ±0.38%, a difference of only 0.03 percentage points. These data demonstrate that this method significantly reduces the number of sub-models and training costs without substantially sacrificing weighing accuracy, thus significantly improving the method's engineering practicality.

[0056] Based on the above implementation method, geometrically invariant characteristic parameters are calculated solely based on the geometric coordinates of the weighing sensor installation locations, neglecting the differences in local stiffness at the bottom of the weighing hopper and the mechanical coupling effect between sensors through the hopper structure. In actual engineering, factors such as the thickness distribution of the weighing hopper bottom plate, the arrangement of reinforcing ribs, and welding deformation can lead to different local stiffnesses at different installation points. Even if two sensors are symmetrically positioned, their output responses may differ significantly when the same load is applied. If mechanical equivalence is judged solely based on geometric coordinates, combinations of working conditions with large differences in mechanical response will be incorrectly classified into the same equivalent group, resulting in a decrease in the weighing accuracy of the shared sub-model.

[0057] To address the aforementioned issues, this embodiment further provides the following specific solution based on the above implementation method. During the system calibration phase, the controller sequentially applies a unit calibration load to each weighing sensor mounting point. Specifically, for the j-th mounting point, the controller controls a standard weight lifting mechanism to vertically apply a standard weight of known mass above that mounting point, while simultaneously recording the output voltage response values ​​of all N weighing sensors. This process is repeated three times at each mounting point, and the average value is taken to construct an N×N dimensional unit load influence coefficient matrix H, where the matrix element h... ij This represents the output response value of the i-th sensor when a load is applied at the j-th mounting point. This matrix comprehensively reflects the true force transmission characteristics of the metering bucket structure, including local stiffness differences and the coupling relationships between the sensors.

[0058] For any normal sensor combination operating condition, assuming the number of remaining normal weighing sensors in this combination is M, the controller extracts an M×M dimensional submatrix H from the unit load influence coefficient matrix H, consisting of the rows and columns corresponding to these remaining normal sensors. sub Specifically, if the remaining sensors are numbered i1, i2, ..., iM Then the rows and columns of H_sub are both taken from the i1, i2, ..., i-th elements in H. M Rows and columns. This submatrix describes the output response of each remaining sensor and the coupling response between them when a load is applied to each remaining sensor mounting point in the current sensor layout.

[0059] The controller operates on an M×M dimensional submatrix H sub Perform singular value decomposition to obtain M singular values, which are then sorted in descending order to form a vector [σ1, σ2, ..., σ]. M Dividing each of these M singular values ​​by M yields the normalized singular value vector [σ1 / M, σ2 / M, …, σ]. M [ / M] is used as the geometrically invariant characteristic parameter for this combined operating condition. This vector comprehensively characterizes the force transmission gain, anisotropy, and local stiffness differences under this sensor layout from the perspective of system energy. The closer the normalized singular value vectors of the two combined operating conditions are, the more equivalent they are in terms of mechanical response.

[0060] After adopting this scheme, tests were conducted on two combined working conditions with symmetrical positions but different local stiffnesses at a six-sensor concrete mixing plant. Based solely on geometric coordinates, these two combinations were determined to be mechanically equivalent and assigned to the same shared sub-model. Using this shared sub-model, 1000 kg of material was weighed for each combination, with weighing errors of ±0.38% and ±0.89%, respectively, the latter showing significantly lower accuracy. However, using the singular value decomposition method of the unit load influence coefficient matrix of this invention, the normalized singular value vectors for the two combinations were calculated as [0.52, 0.28, 0.15, 0.05] and [0.48, 0.25, 0.18, 0.09], respectively. The weighted Euclidean distance between the two was 0.18, exceeding the preset equivalence distance threshold of 0.10, therefore they were determined to be inequivalent and assigned to different mechanically equivalent groups, each training its own independent shared sub-model. After retesting, the weighing errors for the two combinations were ±0.39% and ±0.41%, respectively, showing essentially the same accuracy and a significant improvement. The above data shows that this method can accurately reflect the local stiffness and mechanical coupling characteristics of the measuring hopper, avoid merging errors caused by mechanical inequivalence due to geometric symmetry, and ensure the weighing accuracy of the shared sub-model under various merging conditions while reducing the number of sub-models.

[0061] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Further modifications can be readily implemented by those skilled in the art.

Claims

1. A method for self-diagnosis and isolation of sensor faults in a concrete mixing plant weighing system, characterized in that, Includes the following steps: S1: Connect an independent analog-to-digital converter to the output of each weighing sensor in the concrete mixing plant to convert the analog voltage signal output by each weighing sensor into a digital signal and input it into the controller respectively; S2: A reference sensor is installed in the weighing system near the point of impact of the material falling. The output of the reference sensor is connected to the controller via an independent analog-to-digital converter. S3: During the dynamic process of material falling from the silo to the metering hopper, the controller synchronously captures the transient response waveforms of each weighing sensor at the moment the material hits the metering hopper, and extracts the peak characteristics and decay rate characteristics of the transient response waveforms. S4: The controller compares the peak characteristics of the current transient response waveform of the same weighing sensor with the peak characteristics of the historical transient response waveform of the same weighing sensor. If the deviation exceeds the first preset deviation threshold, a longitudinal peak abnormality is recorded. The controller also compares the decay rate characteristics of the current transient response waveform of the same weighing sensor with the decay rate characteristics of the historical transient response waveform of the same weighing sensor. If the deviation exceeds the third preset deviation threshold, a longitudinal decay abnormality is recorded. The controller compares the peak characteristics of the current transient response waveform of a certain weighing sensor with the average peak value of the current transient response waveform of other weighing sensors and the peak value of the current transient response waveform of a reference sensor. If the deviation exceeds the second preset deviation threshold, a lateral abnormality is recorded. When longitudinal peak abnormality, longitudinal attenuation abnormality and lateral abnormality all occur, the weighing sensor is determined to be faulty. S5: An electronic switch is connected in series in the signal transmission path between each weighing sensor and its corresponding analog-to-digital converter. When the controller determines that a weighing sensor has failed, it sends a disconnect command to the electronic switch corresponding to that weighing sensor to hardware isolate the faulty sensor from the weighing system. S6: The controller uses a pre-trained radial basis function neural network to calculate the total weight of the material based on the output signals of the remaining normal weighing sensors.

2. The method for self-diagnosis and isolation of sensor faults in the weighing system of a concrete mixing plant according to claim 1, characterized in that, In step S3, the controller synchronously captures the transient response waveforms of each weighing sensor at the instant the material impacts the weighing hopper, specifically including the following steps: S301: The controller uses the voltage change rate detected by the reference sensor as the synchronous trigger source. When the voltage change rate exceeds the preset slope trigger threshold, the controller simultaneously sends a trigger sampling command to the analog-to-digital converters corresponding to all weighing sensors. S302: The analog-to-digital converter corresponding to each weighing sensor responds to the trigger sampling command and synchronously acquires the voltage signal sequence of the weighing sensor during the impact period at the same sampling frequency; S303: The controller performs DC bias removal and filtering on the voltage signal sequence collected by each weighing sensor to obtain the original transient response waveform of each weighing sensor. S304: The controller performs amplitude normalization processing on the original transient response waveforms of each weighing sensor. Amplitude normalization processing is to divide the voltage amplitude of each sampling point in the original transient response waveform of each weighing sensor by the reference impulse response peak value recorded by the weighing sensor during the system calibration phase. S305: The controller extracts peak characteristics and decay rate characteristics from the transient response waveform after amplitude normalization.

3. The method for self-diagnosis and isolation of sensor faults in the weighing system of a concrete mixing plant according to claim 2, characterized in that, In step S303, the controller performs DC bias removal and filtering on the voltage signal sequence acquired by each weighing sensor to obtain the original transient response waveform of each weighing sensor. This specifically includes the following steps: S3031: When the material is not falling, the controller continuously records the voltage signal sequence of each weighing sensor and the reference sensor within the second monitoring time window, calculates the average amplitude of the voltage signal sequence of each weighing sensor and the reference sensor within the second monitoring time window, and uses the average amplitude as the dynamic DC bias value of the corresponding weighing sensor and the reference sensor in the current impact event. S3032: The controller subtracts the dynamic DC bias value corresponding to the weighing sensor from the voltage amplitude of each sampling point in the voltage signal sequence collected by each weighing sensor during the impact period to obtain the voltage signal sequence after DC bias removal. S3033: The controller uses the voltage signal sequence of the reference sensor after DC bias removal as the reference input signal of the adaptive filter, and uses the voltage signal sequence of each weighing sensor after DC bias removal as the desired response signal of the adaptive filter. S3034: The adaptive filter iteratively updates the filter weight coefficients based on the error signal between the reference input signal and the desired response signal until the mean square value of the error signal converges to a preset range. The updated filter weight coefficients are then applied to the filtered output signal generated by the reference input signal, which serves as the original transient response waveform of the weighing sensor.

4. The self-diagnosis and isolation method for sensor faults in the weighing system of a concrete mixing plant according to claim 3, characterized in that, In step S305, the controller extracts peak characteristics and decay rate characteristics from the transient response waveform after amplitude normalization, specifically including the following steps: S3051: The controller performs Hilbert transform on the transient response waveform data after amplitude normalization to obtain the envelope curve of the transient response waveform; S3052: The controller searches for the maximum amplitude point in the envelope curve and uses the normalized voltage amplitude corresponding to the maximum amplitude point as the peak characteristic of the transient response waveform; S3053: The controller takes the sampling time corresponding to the maximum amplitude as the starting point, and extracts the attenuation segment data of the envelope curve along the time axis. It then uses the least squares method to fit the attenuation segment data of the envelope curve to an exponential function. The fitting model expression is y(t)=A▪e (-t / τ) , where y(t) is the amplitude of the envelope curve at time t, A is the initial amplitude parameter for fitting, and τ is the decay time constant parameter; S3054: The controller extracts the fitted decay time constant parameter τ as the decay rate characteristic of the transient response waveform.

5. The self-diagnosis and isolation method for sensor faults in the weighing system of a concrete mixing plant according to claim 4, characterized in that, In step S4, the specific steps for the controller to perform feature comparison include: S401: The controller stores the operating condition parameter group corresponding to each impact event and the peak value and decay rate characteristics extracted by each weighing sensor in that impact event. The operating condition parameter group includes the material type identifier, hopper opening parameter and the initial material weight value of the metering hopper. S402: For the current impact event, the controller retrieves the K stored historical impact events based on the current operating condition parameter set, where K is an odd number, and the Euclidean distance between the operating condition parameter set of the K historical impact events and the current operating condition parameter set is minimized. S403: The controller uses the weighted average of the peak characteristics in the K historical impact events as the adaptive peak reference of the weighing sensor under the current impact event, and the weighted average of the attenuation rate characteristics in the K historical impact events as the adaptive attenuation rate reference of the weighing sensor under the current impact event, wherein the weighting coefficient is inversely proportional to the Euclidean distance. S404: The controller calculates the deviation between the current peak characteristic of the weighing sensor and the adaptive peak reference. If the deviation exceeds the first preset deviation threshold, a longitudinal peak abnormality is recorded. The controller also calculates the deviation between the current attenuation rate characteristic of the weighing sensor and the adaptive attenuation rate reference. If the deviation exceeds the third preset deviation threshold, a longitudinal attenuation abnormality is recorded. S405: The controller calculates the average peak value of the current peak characteristics of the other normal weighing sensors in the current impact event, excluding the sensor being evaluated, calculates the ratio of the average peak value to the current peak characteristic of the reference sensor, uses the ratio as the lateral threshold scaling factor, and multiplies it with the second preset deviation threshold to obtain the adaptive lateral deviation threshold. S406: The controller calculates the deviation between the current peak characteristic of the weighing sensor and the average current peak characteristic of other normal weighing sensors. If the deviation exceeds the adaptive lateral deviation threshold, a lateral anomaly is recorded. S407: The controller confirms that the weighing sensor has malfunctioned if it has already recorded a longitudinal peak abnormality based on the longitudinal peak deviation exceeding the first preset deviation threshold, a longitudinal attenuation abnormality based on the longitudinal attenuation deviation exceeding the third preset deviation threshold, and a lateral abnormality based on the lateral peak deviation exceeding the adaptive lateral deviation threshold. S408: If only one or two deviations exceed the corresponding threshold, the controller will maintain the working state of the weighing sensor and trigger an early warning signal, while increasing the sampling frequency and comparison accuracy of subsequent impact events.

6. The self-diagnosis and isolation method for sensor faults in the weighing system of a concrete mixing plant according to claim 5, characterized in that, In step S6, the controller calculates the total weight of the material using a pre-trained radial basis function neural network, specifically including: S601: The controller pre-establishes radial basis function neural network sub-models corresponding to each combination of working conditions, including the combination of working conditions where all weighing sensors in the weighing system are normal and the combination of working conditions where all normal weighing sensors remain after at least one weighing sensor is isolated. The number of input nodes in each sub-model is equal to the number of normal weighing sensors in the corresponding combination of working conditions. S602: The controller uses the sample set consisting of the known material weight and the corresponding steady-state output voltage signal of the normal weighing sensor collected under each combined working condition to train the corresponding radial basis function neural network sub-model, and stores each trained sub-model as a model library with the combined working condition as the index. S603: During online weighing, after the controller completes hardware isolation of the faulty sensor, the controller identifies the combined working conditions composed of the remaining normal weighing sensors and retrieves the radial basis function neural network sub-model that matches the combined working conditions from the model library. S604: The controller inputs the steady-state voltage output signal of the remaining normal weighing sensor to the retrieved sub-model, and uses the output value of the sub-model as the total weight value of the material.

7. The method for self-diagnosis and isolation of sensor faults in the weighing system of a concrete mixing plant according to claim 6, characterized in that, In step S601, the controller pre-establishes radial basis function neural network sub-models corresponding to each normal sensor combination condition, specifically including: S6011: The controller calculates the geometrically invariant characteristic parameters of the remaining normal weighing sensor position distribution under each normal sensor combination working condition based on the installation position coordinates of each weighing sensor at the bottom of the weighing hopper. S6012: The controller determines different combinations of working conditions where the weighted Euclidean distance between geometrically invariant characteristic parameters is less than a preset equivalent distance threshold as mechanically equivalent combinations of working conditions, and divides all normal sensor combinations of working conditions into several mechanically equivalent groups. S6013: The controller trains a radial basis function neural network sub-model for each mechanical equivalence group. This sub-model is used as the sub-model corresponding to each combination of working conditions within the mechanical equivalence group. The sub-model uses the union sample set of known material weights collected under all combination working conditions within the mechanical equivalence group and the steady-state output voltage signal of the corresponding normal weighing sensor as training data, and establishes a mapping relationship from each combination of working conditions to the corresponding sub-model of the mechanical equivalence group.

8. The method for self-diagnosis and isolation of sensor faults in the weighing system of a concrete mixing plant according to claim 7, characterized in that, In step S6011, the calculation of the geometrically invariant characteristic parameters specifically includes: S60111: During the system calibration phase, the controller sequentially applies a unit calibration load to each weighing sensor mounting point, records the output voltage response of each weighing sensor, and constructs an N×N dimensional unit load influence coefficient matrix, where N is the total number of weighing sensors, and the matrix elements represent the output response value of the i-th sensor when the load is applied at the j-th mounting point. S60112: For any normal sensor combination working condition, the controller extracts an M×M dimensional submatrix from the unit load influence coefficient matrix, which is composed of the rows and columns corresponding to the remaining normal weighing sensors in the combined working condition, where M is the number of remaining normal weighing sensors in the combined working condition. S60113: The controller performs singular value decomposition on the M×M dimensional submatrix, sorts the resulting M singular values ​​in descending order and divides them by M to form the normalized singular value vector of the combined working condition, and uses the normalized singular value vector as the geometrically invariant characteristic parameter.